Young, Restless and Creative:Openness to Disruption and Creative Innovations∗

Daron Acemoglu† Ufuk Akcigit‡ Murat Alp Celik§

February 1, 2014

Abstract

This paper argues that openness to new, unconventional and disruptive ideas has a first-order impact on creative innovations– innovations that break new ground in terms of knowledgecreation. After presenting a motivating model focusing on the choice between incremental andradical innovation, and on how managers of different ages and human capital are sorted acrossdifferent types of firms, we provide cross-country, firm-level and patent-level evidence consistentwith this pattern. Our measures of creative innovations proxy for innovation quality (averagenumber of citations per patent) and creativity (fraction of superstar innovators, the likelihoodof a very high number of citations, and generality of patents). Our main proxy for openness todisruption is manager age. This variable is based on the idea that only companies or societiesopen to such disruption will allow the young to rise up within the hierarchy. Using this proxyat the country, firm or patent level, we present robust evidence that openness to disruption isassociated with more creative innovations.

JEL Codes: O40, O43, O33, P10, P16, Z1.

Keywords: corporate culture, creative destruction, creativity, economic growth, entrepre-neurship, individualism, innovation, openness to disruption.

∗We thank Olga Denislamova, Hyunjin Kim, and Gokhan Oz for excellent research assistance, and Pascual Re-strepo and conference and seminar participants at the NBER Productivity group, CIFAR, Brown University, Univer-sity of Pennsylvania, 2014 AEA Meetings and particularly our discussant Joshua Gans, for helpful suggestions andcomments. We also thank the Bilkent University Economics Department for the great hospitality during this project.Financial support from ARO MURI W911NF-12-1-0509 and from the Toulouse Network on Information Technologyis gratefully acknowledged.†Massachusetts Institute of Technology. Email: daron@mit.edu.‡University of Pennsylvania. Email: uakcigit@econ.upenn.edu.§University of Pennsylvania. Email: celik@sas.upenn.edu.

1 Introduction

This paper investigates the impact of economic and social incentives on “creative innovations,”

which we identify with the most influential, innovative and original patents. Though there are

currently more than half a million patents per year granted by the US Patent and Trademark

Offi ce (USPTO), only a handful of those are truly transformative in terms of their contribution

to society’s knowledge and their impact on the organization of production, and probably only a

small fraction account for the bulk of the value created (e.g., Hall, Jaffe and Trajtenberg, 2001,

and further references discussed below). For example, within the field of drugs and medical in-

ventions, there were 223,452 patents between 1975 and 2001, but the median number of citations

of these patents within the next five years was four. A few patents receive many more citations,

however. One was the patent for “systems and methods for selective electrosurgical treatment of

body structures”by the ArthroCare Corporation (with 50 citations), which has also had a major

impact on the field by improving many existing surgical procedures and devices used, inter alia, in

arthroscopy, neurology, cosmetics, urology, gynecology, and laparoscopy/general surgery. Another

example comes from Amazon’s patent for “method and system for placing a purchase order via a

communications network,”which received 263 citations within five years (while the median number

of citations within this class is five) and has fundamentally altered online businesses.

An idea dating back to Joseph Schumpeter (1934) associates creative innovations and entre-

preneurship not only with economic rewards to this type of transformative idea, but also with the

ability and desire of potential innovators and entrepreneurs to significantly deviate from existing

technologies, practices and rules of organization and society and engage in “disruptive innovations.”

This is natural; as Schumpeter emphasizes, innovation is a deviation from existing, inertial ways of

doing things, and thus relies on “mental freedom”from, or even “rebellion”against, the status quo

(pp. 86-94). Similarly, technologies that will cause the most fundamental “creative destruction”

naturally correspond to, and perhaps are driven by, “deviant”and disruptive behavior. This notion

is pithily captured by an inscription prominently displayed on the walls of Facebook’s headquarters

in Silicon Valley:

“Move fast and break things.”

This perspective suggests that societies and organizations that impose a set of rigidly specified

rules, discourage initiative and deviations from established norms, shun or even ostracize rebellious

behavior, and do not tolerate those that “move fast and break things” will significantly lag be-

hind their more open, “individualistic”or “risk-taking”counterparts in creative innovations– even

though they might still be able to function successfully with existing technologies. In the rest of the

1

paper, we thus refer to this constellation of social and economic incentives as openness to disruption

(short for openness to disruptive innovations, ideas and practices).

At the cross-country level, these expectations are borne out when looking at the relationship

between various measures of creative innovations and several proxies for openness to disruption.

Figure 1 gives a glimpse of the cross-country patterns, which we document further below. This

figure uses the average number of citations per patent filed with the USPTO originating from

different countries between the years 1995 and 2000 as the dependent variable in all three panels

(see Section 3 for more information on data and variables). Because of the importance of the US

market for international businesses, companies located all around the world file their more important

innovations with the USPTO, giving us a decent sample size of patents from 50 countries (though

with some notable differences in the number of patents). The average number of citations per

patent is a proxy for the average quality of innovations, since higher-quality patents tend to get

cited more (below we use several other measures as alternative proxies, which all show similar

patterns). Each panel of Figure 1 uses a different proxy for openness to disruption and depicts

the conditional relationship between the average number of citations per patent and the proxy in

question (after controlling for log GDP per capita, average years of secondary schooling and log

total number of patents in the country). It also shows the weighted regression line.1

The first panel focuses on the “individualism” variable of the Dutch social scientist Geert

Hofstede, who constructed indices for various dimensions of “national cultures.”This variable, based

on Durkheim’s (1933) distinction between collectivism and individualism, measures the extent to

which a society functions by relying on loosely knit social ties and thus permits and condones

individual actions even when they conflict with collective goals and practices, particularly in a

business context.2 The second panel uses Hofstede’s “uncertainty avoidance” index, which is an

inverse proxy for a society’s tendency for risk-taking based in part on ideas from Cyert and March’s

seminal (1963) book.

The third panel uses our own proxy for a society’s openness to disruption: the average age

of (top) managers (e.g., CEO and CFO) in the 25 largest listed companies in the country (when

available). The motivation for this variable is that societies that are open to disruption tend to be

more meritocratic in promoting young talent, even those promoting and implementing disruptive

innovations, while those that discourage individualistic risk-taking attitudes tend to make the young

subservient to the old.3

1This is from a weighted regression the using total number of patents as weights to partially correct for the factthat the number of patents is small and thus our measures are quite noisy for several countries in our sample.

2All of our variables are defined and described in greater detail below.3 Interestingly, in the examples of major innovations mentioned above, these were produced by companies with

unusually young leadership. The average age of top managers at ArthroCare Corporation was 41 at the time, andonly 33 at Amazon (compared to an average age of 54.84 among Compustat companies).

2

The average manager age variable, which we collected from publicly available sources, has

several advantages: as opposed to Hofstede’s data it is not subjective; it closely corresponds to a

specific dimension of openness to disruption, and it is available for US companies, enabling us to

conduct most of our analysis with firm-level and patent-level data within the United States.

In all three panels, there is a fairly strong positive relationship, and the weighted regression

line is statistically significant at 5% or less (see below). It is worth noting that this relationship

does not reflect the potential correlation between our measures of creative innovations and GDP

per capita, human capital, or even total number of patents, each of which is controlled for in the

plots and the corresponding regressions. Though far from conclusive, this evidence shows a notable

pattern in the cross-country data on creative innovations, which, to the best of our knowledge, has

neither been noted nor systematically investigated before.

Motivated by these patterns, we first provide a simple model of firm innovation strategies.

Firms can engage in an incremental innovation by building on their existing leading-edge products

or a radical innovation by combining diverse ideas to generate an improvement in a new area.

We assume that some companies have a comparative advantage in radical innovations (because of

their technology, “type” or “corporate culture”), but in addition, their managers’ skills are also

important for the type of innovation. In particular, young managers have more recently acquired

general skills (or are less beholden to a particular type of product or technology). This enables them

to more effectively use current advances in a range of fields to succeed in radical innovation. In

the model, though incremental innovations also increase productivity, it is the radical innovations

that are the engine of growth. This is because incremental innovations in a particular “technology

cluster”run into diminishing returns (as in Akcigit and Kerr, 2010, or Abrams et al., 2013), while

radical innovations create new technology clusters, which increase productivity directly and also

indirectly by making another series of incremental innovations possible.

Our model predicts a reduced-form relationship between manager age and radical innovation.

But this is not necessarily the causal effect of manager age. Rather, manager age is both an

economically relevant variable and more generally a proxy for openness to disruption. In the

model, this is captured by the fact that there is both sorting of young managers to firms that

are open to radical innovation, and also young managers employed by such firms do contribute to

their radical innovations. The model further clarifies that radical innovations will generate higher

quality patents that are more likely to receive a very high number of citations and tend to be

more general in terms of the range of citations they receive (because they are expanding into new

areas). It further predicts another relationship we investigate empirically: products with higher

sales will encourage even high-type firms and young managers to pursue incremental innovations

(because of Arrow’s (1962) replacement effect), and those with many patents will tilt things in

3

favor of radical innovations (because of diminishing returns and more generally because there is a

substantial knowledge base to build upon for such an expansion).

Our model further suggests that institutions or attitudes that ban or discourage expansion into

new areas or combinations that have not been previously experimented with can be highly detri-

mental to radical innovations. Equally, those that prevent young managers from leading companies

could slow down creative innovations by failing to use the more recent skills of such managers that

are necessary for radical innovations. Such institutions and attitudes typically vary across coun-

tries, and this reasoning suggests that similar relationships might be found in the cross-country

data, though the firm-level and patent-level results are our main focus.

The bulk of our paper comprises an empirical investigation of the ideas proposed so far and

illustrated by our theoretical model. After describing the datasets and the various measures we

use to proxy for creative innovations in Section 3, we provide a few more details on cross-country

relationships in Section 4. This involves presenting the regression evidence corresponding to Figure

1 and the results with alternative measures of creative innovations that we also use in our firm-level

analysis. In particular, in addition to the average number of citations per patent, we use three other

main measures: the fraction of superstar innovators, which corresponds to the fraction of patents

accruing to an innovator classified as a “superstar” on the basis of the number of citations; tail

innovations, which we measure as the fraction of patents (of a country or company) that are at

the pth percentile of the overall citations distribution (such as the 99th percentile) relative to those

that are at the median, thus capturing the likelihood of receiving a very high number of citations

normalized by the “median”number of citations; and generality index, constructed by Hall, Jaffe

and Trajtenberg (2001), which measures the dispersion of the citations that a patent receives from

different technology classes. Our results in Section 4 show that patterns similar to those shown in

Figure 1 are also present with these three additional quite distinct measures of creative innovations.

Section 5 then turns to our main empirical focus: the firm-level and patent-level analysis of

openness to disruption and creative innovations. We work with the Compustat sample and use the

age of the CEO (or the average age of top management for robustness) as our proxy for openness to

disruption. We find a very robust correlation between this proxy of openness to disruption and all

of our measures of firm-level creative innovation (with or without a variety of firm-level controls).

Even though we have no compelling strategy to identify an exogenous source of variation in CEO

age (or in openness to disruption at the company level), the firm-level correlations we present are

quite robust.

Perhaps surprisingly, we find similar results when we control for firm fixed effects and exploiting

within firm variation, so that when a younger CEO takes charge, innovations (new patent appli-

cations) become more creative. Exploiting patent-level variation, we also estimate the impact of

4

CEO and inventor age on the creativity of innovations. Our results indicate that both matter,

with roughly similar magnitudes. Interestingly, we also find that younger CEOs tend to work with

younger inventors (though CEO age has a fairly precisely estimated impact even after controlling

for inventor age).

We further use the firm-level data to shed light on our model’s prediction that firms with

greater sales should be less willing to encourage new, potentially disruptive ideas, practices and

innovations, while firms that are technologically more advanced, and thus not able to profitably

function without engaging in major innovations, should be more likely to encourage this type of

disruptive innovation. Our firm-level data enable us to investigate this idea by simultaneously

including interactions of average CEO age with (log) sales and (log) number of patents of the firm.

Though the results here are a little less strong than our main findings, they are broadly consistent

with the notion that CEO age interacts negatively with sales and positively with the number of

patents.

Our paper is related to several literatures. There is a growing literature on the impact of

cultural factors and practices on long-run economic development. The distinction between indi-

vidualist and collectivist cultures is deep-rooted in sociology (e.g., Durkheim, 1933) and has been

widely applied within the sociology, anthropology and psychology literatures (e.g., Parsons, 1949,

Kluckhohn and Strodtbeck, 1961, Schwartz, 1994, Triandis, 1995, and Hofstede, 2001). It has been

emphasized within the economics literature by Greif (1994), though we are not aware of any other

studies emphasizing or empirically investigating the impact of “openness to disruption.”Other as-

pects of cultural practices have been emphasized as major determinants of economic developments

by, among others, Tabellini (2008a,b), Fernandez and Fogli (2009), Guiso, Sapienza and Zingales

(2010), and Alesina, Giuliano and Nunn (2011).

Most closely related to our work are recent papers by Gorodnichenko and Roland (2012) and

Fogli and Veldkamp (2012). Gorodnichenko and Roland also draw a link between innovation and

individualism and provide evidence using Hofstede’s individualism data. Despite the similar mo-

tivating questions, the approaches of the two papers are very different. While Gorodnichenko and

Roland look at aggregate measures of productivity, such as TFP or labor productivity, we focus on

creative innovations defined from patent citations data from the USPTO. We therefore first start

with a microeconomic model of how firms choose their innovation strategies and how managers

of different ages endogenously sort across different types of firms. Though we also show results

with Hofstede’s data, this is only to provide motivating evidence. Our main empirical work in-

stead uses the proxy for openness to disruption we have constructed ourselves based on the age

of managers across countries and, more centrally, focuses on firm-level and patent-level analysis

across US companies. Fogli and Veldkamp also use the individualism index in their theoretical

5

and empirical analysis of “individualistic” social networks and the diffusion of new technologies,

but their emphasis is on how new technologies diffuse over different network structures and their

empirical work exploits exposure to different types of diseases to generate cross-country variation

in societal network structures.

Also closely linked is the small literature on age and creativity. Galenson and Weinberg (1999,

2001), Weinberg and Galenson (2005), Jones and Weinberg (2011) and Jones (2010) provide ev-

idence that a variety of innovators and top scientists are more creative early on, but they also

acquire other types of human capital (perhaps generating different types of creativity) later in their

careers. Jones (2009) develops a model in which scientists have to spend more time mastering a

given area and have to work in teams because the existing stock of knowledge is growing and thus

becoming more diffi cult to absorb and use.4

Schumpeter’s (1934) vision of an innovator as creating disruption, partly in response to economic

incentives and partly for psychological motives that lead them to seek challenges and deviate from

norms, is more closely related to our focus. Traces of this approach can also be seen in Adorno

et al.’s (1950) psychological study of authoritarianism, and in McClelland’s (1961) and Winslow

and Solomon’s (1987) approaches to entrepreneurship (see Kirzner, 1997, for a recent survey).

These ideas have been applied in a cross-country context by Shane (1993, 1995), Hofstede (2001),

Schwartz (1994), Schwartz and Bilsky (1990) and others. To the best of our knowledge, no other

work links these ideas to creative innovations, develops a formal theory along the lines of what we

are attempting here, or provides systematic evidence based on firm- or patent-level data.

Our paper also relates to a large literature on innovation and firm dynamics. Although a few

works (including Acs and Audretsch 1987, 1988, Kortum and Lerner 2000, Baumol 2009, Akcigit

and Kerr, 2010, and Acemoglu et al., 2013) emphasize heterogeneity in innovation behavior and

strategy across firms, we are not aware of other papers in this literature that focus on creative

innovations or link this to the incentives and constraints imposed on the behavior of innovators

and managers within companies. Nevertheless, within this literature, several papers also emphasize

the importance of patent quality. In particular, Trajtenberg (1990), Harhoff et al (1999), Shane

and Klock (1997), Sampat and Ziedonis (2004), and Abrams et al. (2013) document a positive

relationship between citations and various measures of private or social value.

Finally, the main ideas here have a resonance with the innovation literature investigating dis-

ruptive innovations, which follows Christensen’s seminal The Innovator’s Dilemma (1997) and

attempts to explain why many companies are unable to maintain their innovativeness following

4Relatedly, Sarada and Tocoian (2013) investigate the impact of the age of the founders of a company on subsequentperformance using Brazilian data, while Azoulay, Manso and Zivin (2011) show the impact of changes in incentivesdriven by large academic awards and grants on creativity, and Azoulay, Zivin and Wang (2010) investigate the impactof the death of a very productive co-author on academic productivity.

6

success. Henderson (2006) discusses the organizational aspects of the innovator’s dilemma. Adner

and Zemsky (2005) investigate the relationship between disruptive innovations and competition,

while King and Tucci (2002) discuss the role of managerial strategies and experience in dealing

with these issues. Our potential answer to the innovator’s dilemma, consistent both with Arrow’s

replacement effect and our interaction results, is that successful firms with higher sales have more to

fear from disruptive innovations and tend to retrench and become less open to new ideas, practices

and innovations.

The rest of the paper is organized as follows. The next section presents our motivating model.

Section 3 describes our data sources and variable construction and provides a few basic descriptive

statistics. Section 4 presents some basic cross-country correlations corroborating the patterns

shown in Figure 1. Section 5 presents our main empirical results, which are based on firm-level

data. Section 6 concludes.

2 Motivating Theory

In this section, we provide a stylized model of radical and incremental innovations to motivate both

the conceptual underpinnings of our approach and some of our empirical strategies.

2.1 Production

We consider a continuous-time economy in which discounted preferences are defined over a unique

final good Y (t). This final good is produced by labor and a continuum of intermediate goods j,

each located along a circle, C, of circumference 1. The production technology takes the followingconstant elasticity of substitution form

Y (t) =1

1− β

(∫Cqj (t)β kj (t)1−β dj

)Lβ, (1)

where kj (t) denotes the quantity and qj (t) the quality (productivity) of intermediate good j used in

final good production at time t, while L is the total amount of production labor, which is supplied

inelastically.

We follow Klette and Kortum (2004) in defining a firm as a collection of leading-edge (best)

technologies. A perfectly enforced patent for each leading-edge quality technology is held by a

firm, which can produce it at constant marginal cost γ in terms of the unique final good. Because

costs and revenues across product lines are independent, a firm will choose price and quantity to

maximize profits on each of its product lines. In doing so, it will face an iso-elastic inverse demand

derived from the profit maximization of the final good sector, which can be written, suppressing

time arguments, as:

pj = Lβqβj k−βj ,∀j ∈ C.

7

The profit-maximization problem of the firm with leading-edge technology for intermediate good j

can then be written as

π (qj) = maxkj≥0

{Lβqβj k

1−βj − γkj

}∀j ∈ C.

The first-order condition of this maximization problem implies a constant markup over marginal

cost, pj = γ/(1− β), and thus

kj =

[(1− β)

γ

] 1β

Lqj . (2)

Equilibrium profits for a product line with technology qj are

π (qj) = β

[(1− β)

γ

] 1−ββ

Lqj

≡ πqj ,

where the second line defines π.

2.2 Managers

In addition to workers, the economy is also populated by managers. Managers enter and exit the

economy following a stationary Poisson birth and death process, so that the measure of managers,

M , and their age distribution is constant over time. We index a manager by her birth date b.

When a manager is born, she acquires the knowledge associated with the average technology in the

period in which she is born, giving her a knowledge base of

qb ≡∫Cqjbdj.

Similarly, we denote the current period’s knowledge stock– current average technology– by qt ≡∫C qjtdj. Managers will be hired by monopolists to manage production and innovation in their

leading-edge products. In equilibrium, managers will be paid a wage wb,t as a function of the

current period’s technology, qt, and their knowledge, qb. We assume that M < 1, which implies

that the measure of managers is less than the measure of product lines in the economy, so some

product lines will not use a manager. This simplifies the analysis by providing a simple boundary

condition for the determination of equilibrium wages of managers. We also assume that M is

not too small, which will ensure that all firms that need a manager for a “radical innovation,”as

described next, are able to hire one (one can take M → 1 without any loss of generality).

2.3 Innovation Dynamics

The productivity of each intermediate product is determined by its location along a quality ladder

in a given product line. In addition, each leading-edge technology gives the firm an opportunity

for further innovation. Two types of innovations are possible:

8

1. incremental innovations, which improve the productivity of a product line within the current

technology cluster.5 Technology cluster here refers to a specific family of technologies for that

product line. Because incremental innovations take place within this technology cluster, they

will run into diminishing returns. We model this by assuming that the additional produc-

tivity improvements generated by an innovation declines in the number of prior incremental

innovations within a technology cluster. In addition, again because these take place within

a given technology cluster, they build on a narrow technology base and create improvements

over this base. This implies that, as illustrated in Example 1 below, incremental innovations

will have few citations and limited “generality”(captured by the dispersion of citations they

receive from different technology classes as we discuss further below).

2. radical innovations, which combine the current technology of the product line, the knowledge

base of the manager and the available knowledge stock of the economy, to innovate in a new

area (creatively destroying the leading-edge technology of some other firm). This combination

of knowledge creates a new technology cluster (thus akin to Weitzman’s (1998) recombination

approach) upon which new incremental innovations can be built. Because they create new

technology clusters, radical innovations tend to receive more citations, are more likely to have

a very high number of (“tail”) citations, and have greater generality.

Firms can successfully innovate incrementally at the exogenous rate ξ > 0. The nth incremental

innovation in a technology cluster improves the current productivity of product line j by a step

size ηn(qj , qt), where qj is the current productivity of the technology, and qt is the current period’s

technology, and

ηn(qj , qt) = [κqt + (1− κ) qj ] ηαn (3)

with α ∈ (0, 1), η > 0, and κ ∈ (0, 1). This functional form implies two features. First, each innova-

tion builds both on the current productivity of the product line, with weight 1−κ, and on averagetechnology, qt, with weight κ. When this will cause no confusion, we will suppress the arguments

of ηn(qj , qt). Second, productivity gains from incremental innovations decline geometrically, at the

rate α, in the number of prior incremental innovations in the technology cluster.

Denoting by tn the time of the nth incremental improvement for product line j, the evolution

of the technology of product line j in a technology cluster that started with productivity q0j after

n incremental innovations can be written as

qnj = q0j +

∑n

i=0

[κq (tn) + (1− κ) q0

j

]ηαn

= q0j

[1 + (1− κ) η

1− αn+1

1− α

]+ ηκ

∑n

i=0αnq (tn) .

5Our modeling of technology clusters follows Akcigit and Kerr (2010) and Abrams et al (2013).

9

If, instead, there is a radical innovation in this particular product line, the innovator will initiate

a new technology cluster in a different product line (and will still keep its original product line).

First, this ensures a larger improvement on current technology. Second, it generates the ability

to start a new series of incremental innovations. However, radical innovations are not directed,

and since each firm controls an infinitesimal fraction of all products, the likelihood that it will

be the firm itself radically innovating over its own product is zero.6 Thus radical innovations

are associated with “Schumpeterian creative destruction.”We describe the technology for radical

innovations below.

For each of their active product lines, firms hire managers who influence their revenues in two

distinct ways. First, a manager of age a = t− b contributes qtf (a) to the revenues of a firm when

the aggregate technology level is qt (e.g., by reducing costs).7 We presume (but do not need to

impose) that f is increasing, so that more experienced managers are better at cost reductions. If

the firm hires no manager, then it does not receive this additional revenue.

Second, a manager affects the flow rate of radical innovations for firms attempting such radical

innovations, as we describe next.

We assume that there are two types of firms, denoted by θ ∈ {θH , θL} where θH > θL = 0,

which means that firms are distinguished by their “corporate culture”determining their openness

to disruption and radical innovation. We assume that the arrival rate of a radical innovation for a

firm of type θ with a manager with knowledge base qb when the current technology in the economy

is qt is given by

Λθqa, (4)

where

qa ≡ qbqt

is the relative average quality of managers of age a, and Λ ∈ (0, 1] (and the superscript, rather than

the subscript, here indicates that this is a ratio of two averages). This specification implies that

low-type firms, with θL = 0, cannot engage in radical innovations.

Since both high- and low-type firms have the same rate of success, at the rate ξ, when they

attempt incremental innovations, our model also implies that θ captures the comparative advantage

of firms for radical innovation. In addition, young managers also have a comparative advantage in

radical innovation– since the contribution of the manager of age a to cost reductions is the same

6 It may be more plausible to assume that radical innovations also take place over a range of products that are“technologically close”to the knowledge base of the innovator. Provided that there is a continuum of products withinthis range, this would not affect any of our results.

7We model this as an additive element in the revenues of the firm so as not to affect the monopoly price andquantity choices of the firm via this channel.

10

for all firms, and younger managers contribute to the flow rate of radical innovation with high-type

firms.

The parameter Λ captures the role of institutional or social sanctions on radical innovations.

Such sanctions may permit only the implementation of certain radical innovations, thus making

successful innovations less likely.8

Recall also that a successful radical innovation leads to an improvement over the product line

uniformly located on the circle C, and thus generates creative destruction. In particular, if there isa successful radical innovation over a product line with technology qj , this leads to the creation of

a new leading-edge technology (now under the control of the innovating firm and manager), with

productivity

q0j = (1 + η0) qj ,

where the superscript 0 denotes the fact that a radical innovation initiates a new cluster with no

prior incremental innovations.

The next example provides more details on the evolution of technology clusters and the citation

pattern for the patents related to the incremental and radical innovations therein.

Example 1 The following chart provides an illustrative example focusing on two product lines:

First product line:|||

η0︸︷︷︸P1

η1︸︷︷︸P2

η2︸︷︷︸P6

η3︸︷︷︸P7︸ ︷︷ ︸

Tech Cluster 1

|||

η0︸︷︷︸P11

η1︸︷︷︸P12︸ ︷︷ ︸

Tech Cluster 2

|||

η0︸︷︷︸P13

η1︸︷︷︸P14

η2︸︷︷︸P15︸ ︷︷ ︸

Tech Cluster 3

Second product line:|||

η0︸︷︷︸P3

η1︸︷︷︸P4

η2︸︷︷︸P5︸ ︷︷ ︸

Tech Cluster 1

|||

η0︸︷︷︸P8

η1︸︷︷︸P9

η2︸︷︷︸P10︸ ︷︷ ︸

Tech Cluster 2

In this example, Pn denotes the nth patent registered at the patent offi ce and ηn denotes the step

size as described in equation (3). The first technology cluster starts with a radical innovation

associated with a patent P1. The productivity improvement due to this patent is η0. Subsequently

a new incremental innovation in this technology cluster, with patent P2, follows on P1, increasing

productivity by another η1 < η0. After this innovation, there is a radical innovation P3 in the

second product line, followed by two subsequent incremental innovations P4 and P5. Since P5 and

P6 are second incremental innovations in their technology clusters, they increase productivity by

8 In particular, in the context of our modeling of product lines along the circle C, we may assume that such sanctionspermit a firm operating product line j to successfully innovate over technologies that are suffi ciently close to itself.Suppose, for example, that j may be allowed to innovate only on product lines that are at most a distance Λ fromitself. Then the case of no restrictions would correspond to Λ = 1/2, so that radical innovations over any productlines on the circle C are possible, while Λ < 1/2 would correspond to restrictions and thus lower the likelihood ofsuccessful radical innovations.

11

η2 < η1. Note that P1, P3, P8, P11 and P13 are radical innovations starting new technology clusters.

As described above, these come from innovations in other product lines operated by high-type

firms. Suppose also that the firm operating technology cluster 1 with patent P7 is a high-type firm,

and successfully undertakes a radical innovation after P7, launching a new technology cluster on a

different product line, shown above as patent P8, initiating a new technology cluster.

Consider next the patterns of citation resulting from these innovations. It is natural to assume

that each incremental innovation will cite all previous innovations in its technology cluster, which

is the pattern shown in the next table. (Alternatively, such patents might also cite patterns from

previous technology clusters on the same product line, with very similar patterns). In addition, it

is also plausible that, because a radical innovation is recombining ideas from its own product line

and the product line on which it is building, it should be citing the fundamental ideas encapsulated

in the patents that initiated the two technology clusters. For this reason, patents P8, P11, and P13

cite the patents initiating the previous technology cluster in this product line as well as the patent

initiating the most recent technology cluster in their own product line. The next table shows this

citation pattern for the first five patents.

Cited CitingP1 : P2, P6, P7, P8, P11

P2 : P6, P7

P3 : P4, P5, P8

P4 : P5

P5 : none

For example, P2 builds only on P1 and thus only cites P1, and is in turn cited by P6 and P7. P1 is

cited not only by the patents that build on itself within the same product line, P2, P6, P7 and P11,

but also by P8 because this new patent comes out of recombining ideas based on this technology

cluster and those in some other product line. This pattern then implies that radical innovations

will receive more citations and will also receive more “general”citations. They will also be heavily

overrepresented among “tail innovations,”meaning among patents receiving the highest number of

citations. These are the patterns we will explore in our empirical work.

2.4 Entry, Exit and Firm Dynamics

New firms enter at the exogenous flow rate x, and innovate (improve) over an existing product line

uniformly at random. Thus, new firms, upon successful entry, initiate a new technology cluster.

Subsequently, a firm’s type is determined and remains fixed. We assume that with probability

ζ ∈ (0, 1), each becomes a high type, θH = θ, and with the remaining probability, a low type,

θL = 0.

12

As noted above, a firm makes the innovation decision in each of its product lines to maximize

its present discounted value, which we denote by Ws (−→qf ,−→nf ) where s ∈ {H,L}, −→qf is the vectorof productivities of the firm, −→nf is the vector of the number of incremental innovations in each ofthese product lines, i.e.,

−→qf ≡{qf,j1 , qf,j2 , ..., qf,jmf

}, and −→nf ≡

{nf,j1 , nf,j2 , ..., nf,jmf

},

and mf denotes the number of product lines that firm f is operating.9 The value function for a

low-type firm can be written as

rWL (−→qf ,−→nf )−·WL (−→qf ,−→nf ) =

mf∑m=1

maxa≥0 {πqf,jm + qtf (a)− wa,t}

+ξ

[WL

( −→qf \ {qf,jm} ∪ {qf,jm + ηnf,jm+1

},

−→nf\ {nf,jm} ∪ {nf,jm + 1}

)−WL (−→qf ,−→nf )

]−τ [WL (−→qf \ {qf,jm} ,−→nf\ {nf,jm})−WL (−→qf ,−→nf )]

.(5)

We can explain the right-hand side of this value function as follows: for each product line m =

1, ...,mf , the firm receives a revenue stream of πqf,jm as a function of its productivity in this

product line, qf,jm . In addition, it has a choice of the age of the manager it will hire to operate

this product line, and if the manager’s age is a, it will have additional revenue/cost savings of

qtf (a) and pay the market price for such a manager of age a at time t, wa,t. Summing over all

of its product lines gives the current revenues of the firm. In addition, the firm can undertake an

innovation on the basis of each of its active product lines. Since we are looking at a low-type firm,

all innovations will be incremental, thus arriving at the rate ξ. When such an innovation happens

in product line m that has already undergone nf,jm incremental innovations, the mth element of−→qf changes from qf,jm to qf,jm + ηnf,jm+1 and n goes up by one, which we write as the arguments

of the value function changing to −→qf \ {qf,jm} ∪{qf,jm + ηnf,jm+1

}, −→nf\ {nf,jm} ∪ {nf,jm + 1} (and

the firm relinquishes its current value function WL (−→qf ,−→nf )). Finally, the firm might also lose

one of its currently active product lines to creative destruction, which happens at the endogenous

rate τ (which will be determined below), and in that case, the firm’s value function changes from

WL (−→qf ,−→nf ) to WL (−→qf \ {qf,jm} ,−→nf\ {nf,jm}) (i.e., −→qf changes −→qf \ {qf,jm} and −→nf to −→nf\ {nf,jm}).Note also that in writing this value function, we have simplified the notation with a slight

abuse. First, even though the value function depends on calendar time because of its dependence

on average technology in the economy, qt, we have suppressed time as an argument, and second,

we wrote a ≥ 0 instead of a ∈ R+ ∪ {∅} to designate the possibility that the firm may end up not

hiring a manager.

The value function of a high-type firm can be similarly written as

9Here and elsewhere, we suppress time as an explicit argument of the value functions to simplify notation.

13

rWH (−→qf ,−→nf )−·WH (−→qf ,−→nf ) (6)

=

mf∑m=1

max

πqf,jm

+ maxa≥0

qtf (a)− wa,t + ξ

WH

( −→qf \ {qf,jm} ∪ {qf,jm + ηnf,jm+1

},

−→nf\ {nf,jm} ∪ {nf,jm + 1}

)−WH (−→qf ,−→nf )

;

πqm + maxa≥0

qtf (a) + ΛqaθH

EWH

( −→q f ∪ {qj′ + η0

},

−→n f ∪ {0}

)−WH (−→q f ,−→n f )

− wa,t

−τ [WH (−→qf \ {qf,jm} ,−→nf\ {nf,jm})−WH (−→qf ,−→nf )]

.

The intuition for this value function is identical to (5) except for the possibility of a radical innova-

tion. In particular, for each product line m, this high-type firm has a choice between incremental

and radical innovation, represented by the outer maximization. The first option here is choosing

incremental innovation for product line m and is thus similar to the first line of (5). The second

option is radical innovation, and in this case the trade-off involved in the age of the manager is dif-

ferent, since manager age affects the arrival rate of radical innovations as shown in (4). In the case

of a successful radical innovation, the value of the firm changes to EWH

(−→qf ∪ {qj′ + η0

},−→nf ∪ {0}

),

where the expectation is over a product line drawn uniformly at random upon which the radical

innovation will build.

The next proposition shows that these value functions can be decomposed into sums of value

functions defined at the product-line level.

Proposition 1 The value functions in (5) and (6) can be written as

Ws (−→qf ,−→nf ) =

mf∑m=1

Vs (qj , n) ,

where Vs (qj , n) is the (franchise) value of a product line of productivity qj with n incremental

innovations that belongs to a firm of type s ∈ {H,L} such that

rVL (qj , n)−VL (qj , n) = maxa≥0{πqj + qtf (a)− wa,t}+ξ

[VL(qj + ηn+1, n+ 1

)− VL (qj , n)

]−τVL (qj , n) ,

(7)

and

rVH (qj , n)− VH (qj , n)

= max

πqj + maxa≥0

{qtf (a)− wa,t + ξ

[VH(qj + ηn+1, n+ 1

)−VH (qj , n)

]};

πqj + maxa≥0 {qtf (a) + ΛqaθHEVH (qt)− wa,t}

− τVH (qj , n) ,(8)

where EVH (qt) denotes the expected value of a radical innovation when the aggregate technology

level is qt.

14

Proof. Both of these value functions can be derived straightforwardly by conjecturing the

above forms and verifying the conjecture.

2.5 Stationary Equilibrium With κ = 1

We now characterize the stationary equilibrium of this economy in the case where κ = 1– so that

all current innovations build on current technology, qt (and thus not on the current productivity

of the existing technology cluster). This assumption considerably simplifies the analysis, and we

return to the general case where κ < 1 below.

A stationary equilibrium is defined as an equilibrium in which aggregate output, Yt, grows at a

constant rate g, and the distribution of product lines between high- and low-type firms and over

the prior number of incremental innovations remains stationary.

As noted above, firms decide the age of the manager to hire for each of the product lines they

are operating and whether to engage in a radical or incremental innovation. Let us first consider

the value of a product line for a low-type firm. From Proposition 1, we can focus on the decisions

and the value function of such a firm at the product line level, and the relevant value function is

given by (7).

Since some firms will not hire managers (asM < 1), all firms not undertaking radical innovations

must be indifferent between hiring and not hiring a manager, which implies that the equilibrium

wage for managers, employed by firms engaged in incremental innovations, satisfies:

wa,t = qtf (a) . (9)

Substituting the equilibrium wage (9) into (7), we obtain a simplified value function for low-type

firms as

rVL (qj , n)− VL (qj , n) = πqj + ξ[VL(qj + qtηα

n+1, n+ 1)− VL (qj , n)

]− τVL (qj , n) .

Solving this value function gives an explicit characterization of the value function of low-type

firms as shown in the next proposition.

Proposition 2 The value function of a product line operated by a low-type firm, (7) takes the

following form

VL (qj , n) = Aqj +Bqtαn (10)

where

B ≡ ξηαA

r − g + τ + ξ (1− α)and A ≡ π

τ + r.

15

Proof. See the Appendix.

The form of the value function in (10) is intuitive. It depends linearly on current productivity,

qj , since this determines the current flow of profits. It also depends on current economy-wide tech-

nology, qt, since all innovations, including incremental ones, build on this. Finally, it is decreasing

in n since a higher n implies that the next incremental innovation will increase productivity by less

(and incremental innovation is the only type of innovation that a low-type firm can undertake).

We next turn to the value of a product line operated by a high-type firm, which differs from

(7) because high-type firms have to decide whether to engage in incremental or radical innovation,

given by (8) above. Because (4) implies that younger managers have comparative advantage in

radical innovation, it follows straightforwardly that there will exist a maximum age a∗ such that

only managers below this age will work in firms attempting radical innovation. Moreover, the

maximization over the age of the manager in (8) implies that such a firm must be indifferent

between hiring any manager younger than a∗. This implies:

qtf (a∗) + Λqa∗θHEVH(qt)− wa∗,t = qtf (a) + ΛqaθHEVH(qt)− wa,t for all a < a∗.

Note that the oldest manager hired for radical innovation earns (from expression (9))

wa∗,t = qtf (a∗) .

Hence

wa,t =

qtf (a) for a > a∗

qtf (a) + ΛθH [qa − qa∗]EVH(qt) for a ≤ a∗. (11)

This wage schedule highlights that, in general, younger or older managers might be paid more

(this will depend on the f function). Younger managers have a comparative advantage in radical

innovation, but older managers might be more productive in operating firms.10

Now substituting for (11) in (8), we obtain a simplified form of the value function of a product

line operated by a high-type firm as

rVH (qj , n)− VH (qj , n) = max

{πqj + ξ

[VH(qj + qtηα

n+1, n+ 1)− VH (qj , n)

];

πqj + Λqa∗θHEVH(qt)

}− τVH (qj , n) .

We next characterize the solution to this value function and also determine the allocation of

managers to different product lines (and to incremental and radical innovations).

Proposition 3 The value function in (8) takes the following form

VH (qj , n) = Aqj + qtB (n) , (12)

10The evidence in Galenson and Weinberg (1999, 2001), Weinberg and Galenson (2005) and Jones and Weinberg(2011) is consistent with the possibility that either younger or older creative workers might be more productive.

16

where

A =π

r + τ,

and B (n) is given by

(r − g + τ) B (n) =

ξ[Aηαn+1 + B (n+ 1)− B (n)

]for n < n∗

Λqa∗θH

[(1 + η) A+ B (0)

]for n ≥ n∗

, (13)

where n∗ ∈ Z++ is the number of incremental innovations within a technology cluster at which there

is a switch to radical innovation given by

n∗ =⌈n′⌉such that ξ

[Aηαn

′+1 + B(n′ + 1

)− B

(n′)]

= Λqa∗θH

[(1 + η) A+ B (0)

]. (14)

Proof. See the Appendix.

The intuition for this high-type value function is similar to that for Proposition 2, except that

the dependence on the number of prior innovations in the current technology cluster, n, is more

complicated since when n exceeds n∗, a high-type firm will switch to radical innovation (and from

that point on n will no longer be relevant). This critical value n∗ is given by (14); intuitively, it

is the smallest integer after n′ where n′ equates the value of attempting an additional incremental

innovation to the value of attempting a radical innovation (the notation dne denotes the next integerafter n).

It is also worth noting that this threshold, n∗, is constant in the stationary equilibrium. This is

because the value function increases linearly in qt, but the knowledge stock and wages of managers

also increase linearly, and in the stationary equilibrium, these two forces balance out, ensuring that

n∗ is constant while VH increases linearly in qt.

Given the form of VH , EVH(qt), the value of a new radical innovation, can be written as

EVH(qt) = E[Aqj + Aηqt + qtB (0)

]= [A (1 + η) + B (0)]qt

≡ vqt,

where the last line defines v. Then the equilibrium wage schedule simplifies to:

wa,t =

f (a) qt for a > a∗

[f (a) + ΛθH(qa − qa∗)v]qt for a ≤ a∗. (15)

and is thus also linear in qt.

Our main results in this subsection fully characterize the stationary equilibrium.

17

Proposition 4 At time t managers with a ≤ a∗ (“young”managers or those with b ≥ b∗t ) will be

hired on product lines for which firms are pursuing radical innovations, which are those operated by

high-type firms and that have had more than n∗ prior incremental innovations, where n∗ is given

by (14). Managers with a > a∗ (“old”managers or those with b < b∗t ) will be hired by firms that

undertake incremental innovations. Managerial wages at time t are given by (15).

A lower Λ (corresponding to the society being less permissive to radical innovations) will increase

n∗ (so that a lower fraction of high-type firms will pursue radical innovation), and will reduce the

wages of young managers (because there is less demand for the knowledge of young managers).

Proof. This result directly follows from Propositions 2 and 3.

The implications of changes in Λ are particularly interesting. A lower value of this parameter

naturally reduces radical innovations and, at the same time, decreases the wages of young managers,

thus making it look like the society is discriminating against the young; but in fact this is a

consequence of the society discouraging radical innovations.

Our empirical work is partly inspired by Proposition 4 (though it does not directly test its

results). As explained above, radical innovations will be associated with greater indices of our

measures of creative innovations (innovation quality, tail innovations, superstar fraction, and gen-

erality). We will first investigate the relationship between society-wide measures of permissiveness

to radical innovations (corresponding to Λ) and our measures of creative innovations in the cross-

country data. We then turn to the relationship in firm- and patent-level data between manager

(CEO) age and creative innovations. In both cases, as our model highlights, the relationship be-

tween manager age and creative innovations does not capture a causal effect of manager age on

creativity or innovation. In our model, this can be seen from the sorting of young managers to

high-type firms, which are the only ones with the ability to undertake radical innovations. But at

the same time, for a high-type firm, a young manager contributes to radical innovations, so that

there is a causal effect of manager age, but this will be confounded with the aforementioned sorting

when we look at the correlation between manager age and measures of radical innovation. For this

reason, we do not interpret the correlations we present below as causal effects, though at the end of

our firm-level analysis we present a tentative strategy for decomposing these correlations between

causal and sorting effects.

2.6 General Equilibrium and the Stationary Distribution of Products

We next characterize the stationary distribution of product lines in the economy in terms of their

prior number of incremental innovations and then use this distribution to determine the aggregate

growth rate of the economy in the stationary equilibrium.

18

The aggregate creative destruction rate in the economy results from entry and radical innova-

tions and can be written as

τ = x+M

∫ a∗

0ΛqaθdF (a) ,

where x is the entry rate, F (a) denotes the stationary distribution of manager age, and a∗ is the

threshold below which managers are hired by firms to do radical innovation. We can further split

this aggregate creative destruction rate into its components coming from low- and high-type firms:

τL = x (1− ζ) and τH = xζ +M

∫ a∗

0ΛqaθHdF (a) .

Clearly τ = τH+τL. Note that low-type firms generate creative destruction only when they initially

enter the economy (since they do not engage in radical innovation).

Let us denote the fraction of product lines occupied by high- and low-type firms with n prior

incremental innovations by, respectively, µHn and µLn (these are not functions of time as we are

focusing on a stationary equilibrium). Naturally,∑∞

n=0

[µHn + µLn

]= 1.

The invariant step size distribution is determined by the following flow equations for high types

Outflow Inflow(τ + ξ)µH0 = τH for n = 0(τ + ξ)µHn = ξµHn−1 for n∗ > n > 0

τµHn∗ = ξµHn∗−1 for n = n∗

µHn = 0 for n > n∗

.

Intuitively, entry into the state of high-tech product lines with n = 0 is driven by radical innovation

from high-type firms, which takes place at the flow rate τH . Exit from this state takes place

when the current firm engages in an incremental innovation, at the flow rate ξ or when there is

creative destruction (from both high- or low-type firms), which takes place at the aggregate creative

destruction rate, τ . Entry and exit into other states have similar intuitions, except that there is no

entry into states with n > n∗ since high-type firms switch to radical innovation at n = n∗.

The flow equations for the low-type product lines can be written similarly as:

Outflow Inflow(τ + ξ)µL0 = τL for n = 0(τ + ξ)µLn = ξµLn−1 for n > 0

.

These can be solved for the following geometric distributions for high- and low-type firms:

µLn =

[ξ

τ + ξ

]n τL

τ + ξfor n ∈ Z+, and µHn =

[

ξτ+ξ

]nτH

τ+ξ for n < n∗[ξ

τ+ξ

]nτH

τ for n = n∗.

19

To derive the aggregate growth rate, let us also combine (1) with (2), which gives

Y =L

1− β

[(1− β)

γ

] 1−ββ

q

The growth rate of the economy is then equal to the growth of the average quality qt. After a small

time interval ∆t > 0, the average quality evolves according to the following law of motion:

qt+∆t = qt + ηqt[x+ µHn∗QΛθ

]∆t+ qtξη∆t

[∑n∗

0µHn α

n +∑∞

0µLnα

n

]+ o(∆t),

where Q ≡ 1F (a∗)

∫ a∗0 qadF (a) is the average productivity gap of the managers that are hired for

radical innovations, and o(∆t) denotes terms that are second order in ∆t. Then the growth rate in

the stationary equilibrium can be obtained as:

g = η[x+ µHn∗QΛθ

]+ ξη

[∑n∗

0µHn α

n +∑∞

0µLnα

n

].

2.7 Equilibrium With κ < 1

In this subsection, we turn to the general case with κ < 1. We will show that the structure of

the equilibrium is similar, except that now the switch to radical innovation for high-type firms will

depend both on their current productivity and on their prior incremental innovations.

The value of a product line operated by low- and high-type firms can now be written, respec-

tively, as:

rVL (qj , n)− VL (qj , n) = maxa≥0{πqj + qtf (a)− wa,t}+ ξ

[VL(qj + ηn+1, n+ 1

)− VL (qj , n)

]− τVL (qj , n) ,

and

rVH (qj , n)− VH (qj , n) = max

πqj + maxa≥0

{qtf (a)− wa,t + ξ

[VH(qj + ηn+1, n+ 1

)−VH (qj , n)

]};

πqj + maxa≥0 {qtf (a) + ΛqaθHEVH(t)− wa,t}

−τVH (qj , n) .

Here note that, with a slight abuse of notation, we wrote EVH(t) instead of EVH(qt) for the value

of a new radical innovation, since this depends in general not just on average current productivity

in the economy, qt, but the distribution of product lines across different states. All the same, in the

stationary equilibrium it will clearly grow at the same rate as qt, g. Second, ηn is now a function of

both the current productivity of the firm and the average current productivity in the economy, qt.

With an argument similar to that in the previous subsection, the equilibrium wage schedule for

managers will be given by

wa,t =

f (a) qt for a > a∗

f (a) qt + ΛθH[qa − qa∗

]EVH(t) for a ≤ a∗

20

This enables us to write simplified versions of the value functions as:

rVL (qj , n)− VL (qj , n) = πqj + ξ[VL(qj + ηn+1, n+ 1

)− VL (qj , n)

]− τVL (qj , n)

rVH (qj , n)− VH (qj , n) = max

{πqj + ξ

[VH(qj + ηn+1, n+ 1

)− VH (qj , n)

];

πqj + Λqa∗θHEVH(t)

}− τVH (qj , n) .

Proposition 5 Consider the economy with κ < 1. Then, for a product line with current quality

q operated by a high-type firm, the manager will be younger and will pursue radical innovation

when the number of prior incremental innovations is greater than or equal to n∗t (q), where n∗t (q)

is increasing in q. That is, a high-type firm is more likely to pursue radical innovation when its

current productivity is lower and the number of its prior innovations in the same cluster is higher.

Proof. See the Appendix.

This proposition thus establishes that in this generalized setup (with κ < 1), radical innovation

is more likely when a high-type firm has lower current productivity (conditional on its prior number

of incremental innovations), or conversely, for a given level of productivity, it is more likely when

there has been a greater number of prior incremental innovations. We will investigate this additional

implication in our firm-level analysis.

3 Data and Variable Construction

In this section, we describe the various datasets we use and our data construction. We also provide

some basic descriptive statistics.

3.1 Data Sources

USPTO Utility Patents Grant Data (PDP) The patent grant data are obtained from the

NBER Patent Database Project (PDP) and contain data for all 3,279,509 utility patents granted

between the years 1976-2006 by the United States Patent and Trademark Offi ce (USPTO). This

dataset contains extensive information on each granted patent, including the unique patent number,

a unique identifier for the assignee, the nationality of the assignee, the technology class, and back-

ward and forward citations in the sample up to 2006. Following a dynamic assignment procedure,

we link this dataset to the Compustat dataset, which we next describe.11

Compustat North American Fundamentals The Compustat data for publicly traded firms

in North America are from Wharton Research Data Services. This dataset contains a detailed list of

balance sheet items reported by the companies annually between 1974 and 2006. It contains 29,378

different companies, and 390,467 company × year observations. The main variables of interest are11Details on the assignment procedure are provided at https://sites.google.com/site/patentdataproject/.

21

net sales, employment, firm age (defined as time since entry into the Compustat sample), SIC code,

R&D expenditures, total liabilities, net income, and plant property and equipment as a proxy for

physical capital.

Executive Compensation Data (Execucomp) Standard and Poor’s Execucomp provides in-

formation on the age of the top executives of a company starting from 1992. We use information

on CEO age or the average age of (top) managers of a company to construct proxies for openness

to disruption at the firm level.12

The Careers and Co-Authorship Networks of U.S. Patent Inventors Extensive informa-

tion on the inventors of patents granted in the United States between years 1975-2008 is obtained

from Lai et. al.’s (2009) dataset. These authors use inventor names and addresses as well as patent

characteristics to generate unique inventor identifiers upon which we heavily draw. Their dataset

contains 8,031,908 observations at the patent × inventor level, and 2,229,219 unique inventors, andcan be linked to the PDP dataset using the unique patent number assigned by the USPTO.

National Culture Dimensions The Dutch social scientist Geert Hofstede devised five different

indices of national culture: power distance, individualism vs. collectivism, masculinity vs. femi-

ninity, uncertainty avoidance, and long-term orientation. The initial survey was conducted among

IBM employees in 30 countries to understand cross-country differences in corporate culture. This

work has been expanded with additional surveys that have been answered by members of other

professions and expanded to 80 countries (see Hofstede, 2001, and Hofstede et. al., 2010).13 We

use the individualism and uncertainty avoidance measures below.

The individualism measure is defined as “a preference for a loosely-knit social framework in

which individuals are expected to take care of themselves and their immediate families only.”A

low individualism score is indicative of a more collectivist society, where social safety networks are

more common and individuals are influenced by collective goals and constraints.

The uncertainty avoidance measure expresses the degree to which the members of a society seek

to avoid uncertainty and ambiguity. Countries with a higher score are more rigid in terms of belief

and behavior and are more intolerant of unorthodox ideas. On the other end of the spectrum,

societies with a low score are more welcoming to new ideas and value practice above principles.

Both the individualism and the uncertainty avoidance indices are normalized to lie between 0 and

1.12We drop of observations where reported CEO age is less than 26.13http://geert-hofstede.com/national-culture.html

22

Cross-Country Data on Manager Age We also collected data on the age of the CEOs and

CFOs of the 25 largest listed companies for 37 countries. We selected the top 25 companies,

when available, according to the Financial Times’FT-500 list, which ranks firms according to their

market capitalization. We completed the list by using information from transnationale.org when

the FT-500 did not include 25 companies for a country. We then obtained the age of the CEOs

and CFOs from the websites of the companies. Overall, our dataset has on average 20 companies

and 31.6 managers (CEO or CFO) per country.

Other Data Sources The average years of schooling in secondary education is used as a proxy

of the human capital of a country, retrieved from the Barro-Lee dataset (Barro, Lee forthcoming)14.

Real GDP per capita numbers and R&D intensity come from the World Bank’s World Development

Indicators database.

In our baseline analysis, we focus on citation and patents between 1995 and 2000 (with patents

classified according to their year of application). This is motivated by our wish to construct a

balanced panel of firms in our baseline firm-level analysis, where we use a single observation per

firm (extending the beginning of our sample to the earliest date at which we have manager age,

1992, significantly reduces our sample). We also stop the sample in 2000 (or 2002), so that we have

a suffi cient subsequent window during which to measure citations. We show below that our results

are robust to extending the sample.

3.2 Variable Construction

Innovation Quality Our baseline measure of innovation quality is the number of citations a

patent received as of 2006. We also use the truncation correction weights devised by Hall, Jaffe,

and Trajtenberg (2001) to correct for systematic citation differences across different technology

classes and also for the fact that earlier patents will have more years during which they can receive

citations (we also experiment with counting citations during a five-year window for each patent).

Based on this variable, an average innovation quality variable is generated at the country × yearand company × year levels. For our cross-country dataset, the country of the assignee is used todetermine the country to which the patent belongs.

Superstar Fraction A superstar inventor is defined as an inventor who surpasses his or her peers

in the quality of patents generated as observed in the sample. A score for each unique inventor

is generated by calculating the average quality of all the patents in which the inventor took part.

14http://www.barrolee.com/data/dataexp.htm

23

All inventors are ranked according to this score, and the top 5% are considered to be superstar

inventors. The superstar fraction of a country or company in a given year is calculated as the

fraction of patents with superstar inventors in that year (if a patent has more than one inventors, it

gets a fractional superstar designation equal to the ratio of superstar inventors to the total number

of inventors on the patent). The country of the inventor is determined according to the country of

the patent assignee.

Tail Innovations The tail innovation index is defined as the fraction of patents of a firm or

country that receive more than a certain number of citations (once again using the truncation

correction weights of Hall, Jaffe and Trajtenberg, 2001). Namely, let sft(p) denote the number

of the patents of a firm (or country) that are above the pth percentile of the year t distribution

according to citations. Then, the tail innovation index is defined as

Tailft(p) =sft(p)

sft(0.50),

where p > 0.50. This is of course also equivalent to the ratio of the number of patents by firm f at

time t with citations above the pth percentile divided by the number of patents by firm f at time

t with citations above the median (and is not defined for firms that have no patents with citations

above the median). For our baseline measure of tail innovations, we choose p = 0.99, so that our

measure is the fraction of patents of a firm or country that are at the 99th percentile of citations

divided by the fraction of patents that are at the median of citations. The reason we include

sft(0.50) in the denominator is that we would like to capture whether, controlling for “average”

innovation output, some countries, companies or innovators have the tendency for generating “tail

innovations”with very high citations.

Generality and Originality We also use the generality and originality indices devised by Hall,

Jaffe and Trajtenberg (2001). Let i ∈ I denote a technology class and sij ∈ [0, 1] denote the share

of citations that patent j receives from patents in technology class i (of course with∑

i∈I sij = 1).

Then for a patent j with positive citations, we define

generalityj = 1−∑i∈I

s2ij .

This index thus measures the dispersion of the citations received by a patent in terms of the

technology classes of citing patents. Greater dispersion of citations is interpreted as a sign of

greater generality. The originality index is defined similarly except that we use the citations it

gives to other patents. Both indices are averaged across all of the patents of a firm or a country to

obtain our firm-level and cross-country originality and generality indices. The patent classes used

are the 80 two-digit International Patent Classification (IPC) classes.

24

3.3 Descriptive Statistics

Panel A of Table 1 provides descriptive statistics for our cross-country, balanced firm and unbal-

anced firm samples. Since we focus on regressions weighted by the number of patents held by a

company or country, all statistics are weighted by the number of patents. We multiply our indices

for tail innovation, superstar fraction, generality, and R&D intensity by 100.

The table shows that average manager age is 56.1 in our cross-country sample and 52.3 in

our firm-level (balanced or unbalanced Compustat) sample, while average CEO age is 55.3 in the

balanced sample and 55.5 in the unbalanced sample. The comparison of our average number

of citations per patent, superstar fraction, tail innovation, and generality indices shows that, as

expected, our Compustat firms have higher values than the average country.

Panel B of Table 1 presents the correlation between our three cross-country indices of openness

to disruption and between our country-level and firm-level measures of creative innovations. These

three indices are quite highly correlated. Panels C and D present the cross-section country and

firm-level correlations between our main measures of creativity of innovations, which are also quite

highly correlated except for the generality index at the firm level.

4 Cross-Country Correlations

We start our empirical analysis by providing a few more details on the cross-country patterns shown

in Figure 1. Though our main evidence comes from firm-level and patent-level regressions presented

in the next section, the results in this section show that the cross-country patterns discussed in the

Introduction are also fairly robust.

We should note at this point that, as already mentioned, the interpretation of cross-country

and firm-level results could in fact be different. At the firm level, as our theory highlighted and we

will emphasize again below, the age of managers is in part an indicator for a certain type of firm or

“corporate culture”(corresponding to our high-type firms) that can undertake radical innovations.

Nevertheless, conditional on having such a firm, a young manager does also contribute to radical

innovations (because of his more recent knowledge stock). At the country level, however, manager

age, like our other measures, is likely to have much of its impact on the creativity of innovations

entirely through institutions, attitudes and values of the society within which it is correlated.

Table 2 reports OLS results from cross-country regressions of the following form:

yc = αIc +X′cβ + εc, (16)

where yc is one of our measures of creative innovations (innovation quality, superstar fraction, tail

innovation, or generality) for country c, Ic denotes one of our measures of openness to disruption

25

(the individualism index, the uncertainty avoidance index, or average manager age),Xc is a vector of

controls (including average log GDP per capita of the country, average years of secondary schooling

and log of total patents of the country during this time period), and finally, εc is an error term. The

coeffi cient of interest is α, which will reveal whether there is a cross-country correlation between

our measures of openness to disruption and the creativity of innovations.

All regressions in Table 2 include one observation per country, report robust (against het-

eroscedasticity) standard errors, and use the total number of patents as weights. The weighted

specification is motivated by the fact that countries with more patents are both more important

for their contribution to creative innovations and have much more precisely estimated measures

for our key variables. Appendix Table A1 shows the distribution of total number of patents across

countries.15

Panel A of Table 2 focuses on Hofstede’s individualism index and contains results for our four

main measures of creative innovations. Column 1, for example, has an estimate of 4.97 (standard

error = 2.46) in the first row for innovation quality (average number of citations). This estimate thus

shows a positive association between the individualism index and the average number of citations

per patent. The other rows show the effect of log GDP per capita, average years of secondary

schooling and log total number of patents. GDP per capita and average years of secondary schooling

are not significant, a pattern common with most other specifications we report in this table, while

log patent count is significant and indicates that countries that have more patents also tend to have

more citations per patent. The quantitative magnitude of the correlation between individualism

and innovation quality is sizable. Moving from the country at the 25th percentile of individualism

in our sample to those at the 75th percentile (from 0.19 to 0.73) increases our measure of innovation

quality by 19% relative to the weighted sample mean (14.5). In fact, the R2 at the bottom of the

panel indicates that, despite its parsimony, this specification explains 73% of the cross-country

variation in average citations per patent.

The other columns of the table show the correlation between the individualism index and our

other measures, in particular the superstar fraction of innovations, tail innovations, and generality.

In each case, there is a fairly strong and highly significant correlation between individualism and

these indices (the estimate of the coeffi cient on individualism is significant at less than 1% in

all of these cases). The magnitudes for the superstar fraction and tail innovation variables are

significantly larger than in the first column: in the former case, moving from the country at the

15An additional covariate that might be useful to control for would be the average educational attainment ofmanagers in a country. Though this number is available in the World Bank dataset that Gennaioli et al. (2013)use, there is very little overlap between this developing country sample and ours. We have instead experimentedwith controlling for the average education of the managers of the companies we have used for compiling our averagemanager age variable. This has no effect on the results reported here and is omitted to save space. The details areavailable upon request from the authors.

26

25th percentile of individualism in our sample to that at the 75th percentile increases the superstar

fraction by 80% relative to the weighted sample mean (6.68), whereas in the latter case, it increases

by 67% relative to the weighted sample mean (1.92).

Panel B has exactly the same structure, except that the right-hand-side variable is Hofstede’s

uncertainty avoidance index. The patterns are very similar and generally even more precisely esti-

mated (though, of course, they are now negative, since greater uncertainty avoidance corresponds

to less openness to disruption). The quantitative magnitudes are also similar to those in Panel A.

In Panel C, we turn to our measure of manager age. This reduces our sample from 50 to

37 due to the more limited availability of this measure. As we argued above, we believe this is

a good proxy for openness to disruption because only companies and countries that are open to

new and potentially disruptive ideas, innovations and practices enable young managers to rise up

to the highest positions rather than relying on age, experience and slow movements within the

hierarchy. The patterns are very similar in this case also, with a strong correlation between average

manager age and all four of our measures of creative innovations. The quantitative magnitudes are

also broadly similar in this case. For example, moving from the country at the 25th percentile of

average manager age in our sample to the 75th percentile (from 51.5 to 54.3) reduces our measure of

innovation quality by 9.4% relative to the sample mean (14.5). More relevant for comparison with

our firm-level and patent-level results might be to look at how much a one-year change in manager

age impacts our various measures of creative innovations. Here the results are again reasonable.

For example, such a change would increase average citations by 0.48 (3.3% compared to its mean

of 14.5), the superstar fraction by 0.96 (14.4% relative to its mean), tail innovations by 0.23 (11.7%

relative to its mean) and generality by 0.28 (1.3% relative to its mean).16

Tables 3 and 4 probe the robustness of the cross-country relationships reported in Figure 1

and Table 2. Table 3 looks at various alternative measures of creative innovations (which we also

investigate at the firm level). These are average citations per patent but now constructed using only

a five-year window (so that we do not have to rely on the correction factors); an alternative measure

of the superstar fraction of patents but now computed using information on the most highly cited

patent to the inventor (rather than lifetime average citations); the tail innovation index computed

with p = 0.90 (instead of p = 0.99); and the originality index mentioned above. The results in all

cases are similar to the baseline (though weaker and not statistically significant with the alternative

measure of superstar fraction).

Table 4, on the other hand, investigates whether these results can be explained by the fact

that R&D intensity (defined as total R&D spending divided by GDP at the country level) differs

16We do not run regressions including multiple indices at the same time, since we believe this type of horse racewould not be particularly informative. Instead, we interpret each of these variables as a proxy for the same underlyingtendency for openness to innovation, new practices and ideas.

27

across countries. Our results largely might be reflecting the fact that some countries invest more

in R&D and as a result generate more creative innovations. However, in our sample R&D intensity

is not systematically related to individualism, uncertainty avoidance, or average manager age.

Moreover, Table 4 shows that controlling for variation in R&D intensity does not change the basic

correlations in our sample. The parameter estimates do change in some cases, particularly with

the individualism variable, but the association between our measures of openness to disruption and

creativity of innovations always remains highly significant.17

5 Firm-Level and Patent-Level Results

Our main empirical results exploit firm-level variation in manager age across Compustat companies.

Though several other factors determine manager age, we believe that, as with the cross-country

variation, a younger manager/CEO reflects a greater openness to disruption. Motivated by this rea-

soning, in this section we show the relationship between firm-level measures of creative innovations

and manager age.

Two caveats are important at this point. First, our theoretical results relate manager age at

the product-line level to the innovation strategy and creativity of innovations, while the bulk of our

empirical analysis in this section will be at the firm level focusing on the age of a firm’s CEO (or

top managers). Second, consistent with our theoretical framework where different types of firms

select into hiring of young managers (recall Propositions 2 and 5), we do not interpret the results

we report in this section as the “causal effect”of manager age on creative innovations. Instead, we

believe that manager age at least partly proxies for the company’s overall openness to disruption.

At the end of this section, we turn to a more direct investigation of the effect of manager age on

creative innovations.

5.1 Main Results

Our main results are presented in Table 5. Our estimating equation is similar to (16),

yf = αmf +X′fβ + δi(f) + εf , (17)

where yf is one of our measures of creative innovations (innovation quality, superstar fraction, tail

innovation, or generality) for firm f , and mf is our firm-level measure of openness to disruption,

the average age of company CEOs over our sample window. In addition, Xf is a vector of controls,

in this case, firm age, log of employment, log of sales, and log of total number of patents during our

17We also experimented with using cross-country differences in demographics to instrument for average managerage differences. Though these results corroborate the patterns shown here, we do not report them both becausedemographics could have a direct effect on the creativity of innovations, invalidating the exclusion restriction of sucha strategy, and because we view the cross-country results as motivation rather than as our main empirical focus.

28

time window (we do not have measures of the human capital of the firm’s employees).18 Controlling

for firm age is particularly important, since we would like to distinguish the correlation of creativity

of innovations with manager age from its correlation with firm age. In addition, δi(f) denotes a full

set of four-digit main SIC dummies included in all regressions so that the comparisons are always

across firms within a fairly narrow industry. Finally, εf is the error term.

Our baseline sample comprises 279 firms with complete information on CEO and positive patents

between 1995 and 2000 (as well as information on firm age, sales, and employment). We first exploit

only cross-sectional information, so our regressions have one observation per firm and are weighted

with the total patent count of the firm. All standard errors are again robust. Different columns of

Table 5 correspond to our four different measures of creative innovations, now constructed at the

firm level.

Column 1 shows an economically sizable correlation between CEO age and our measure of

innovation quality (average number of citations per patent). The coeffi cient estimate, −0.278

(standard error = 0.088), is statistically significant at 1% and indicates that companies with a

younger CEO have greater innovation quality. We interpret this pattern as evidence that companies

that are more open to disruption tend to be the ones producing more creative innovations. The

quantitative magnitudes are sizable and comparable to the quantitative magnitudes we obtained

in the cross-country data when using average manager age. For example, the effect of a one-year

increase in CEO age is to raise average citations by 0.278, which is about 60% of the magnitude of

the cross-country relationship between average manager age and innovation quality.19

The pattern of the covariates is also interesting. Firm age is negatively associated with innova-

tion quality, suggesting that younger firms are more creative (though this pattern is not as robust

as the impact of CEO age in other specifications). Our measures of creative innovations are also

uncorrelated with employment and sales and largely uncorrelated with the number of patents held

by the firm (the exception being a marginally significant relationship for tail innovations). This

confirms that our measures of creativity of innovations are quite distinct from the total number of

patents.

Column 2 shows a similar relationship with the superstar fraction (−0.300, standard error =

0.141). This also suggests that younger CEOs tend to work with higher-quality innovators (a

relationship we directly investigate in Table 10 below). Columns 3 and 4 show even more precisely

estimated (significant at 1% or less) and economically large relationships with our measures of tail

18Our log employment and log sales variables and the cross-sectional regressions are computed as averages of annuallog employment and log sales.19 If, instead, we use the average age of top management, the quantitative impact rises to 0.418; which is more

comparable to the cross-country magnitude. The quantitative implications of moving from the 75th percentile of themanager age distribution to the 25th percentile are also more similar to the cross-country magnitudes.

29

innovations and generality.

Overall, these results suggest that there is a strong statistical and quantitative relationship

between the age of the CEO of a Compustat company and each one of our four measures of

creative innovations. Though this relationship may not be causal (or may reflect the impact of

CEO age working through other channels than openness to disruption), it is both new and quite

consistent with our theoretical expectations. We will next see that it is also quite robust.

5.2 Robustness

Tables 6 and 7 probe the robustness of our firm-level results in different dimensions. Table 6 looks

at the same alternative measures of creative innovations we studied in Table 3 in the cross-country

context (recall that these are a measure of innovation quality using average citations per patent

computed using only five years of citations data, a measure of superstar inventors using information

on the most highly cited patent of the inventor, the tail innovation index with p = 0.90, and the

originality index). The results show that the pattern is quite similar to those in Table 5, except that

the relationship is no longer statistically significant with the alternative measure of the superstar

fraction.

Table 7 looks at several different robustness exercises. Panel A replaces the four-digit SIC

dummies with three-digit dummies, with effects very similar to our baseline results.

Panel B goes in the opposite direction and enriches the set of controls. In particular, this

specification, in addition to the four-digit SIC dummies, includes several other firm-level controls:

profitability (income to sales ratio), debt to sales ratio, and log physical capital of the firm. The

results are virtually the same as those in Table 5, but somewhat more precisely estimated. For

example, CEO age is statistically significant at less than 1% with all of our measures of creative

innovations, except for the superstar fraction, for which it is significant at 5%.

Panel C, in addition, includes R&D intensity (R&D to sales ratio) in the previous specification.20

This is intended, as in the cross-country context, to verify that our results cannot be explained by

some firms performing more R&D than others (here the sample declines to 257 companies). The

results are once again very similar to those in our baseline regressions in Table 5.

Panel D uses the average age of the top management team rather than CEO age. We prefer CEO

age as our baseline measure because across companies there is considerable variation in the number

of managers for which age data are available, making this measure potentially less comparable

across firms. Nevertheless, the relationship is very similar to this measure as shown in Panel D.

Panels E and F reestimate the specifications in Table 5 for subsamples of high-tech and low-tech

firms, where high-tech firms are those in SIC 35 and 36 (industrial and commercial machinery and

20To deal with outliers in R&D expenditures, we winsorize this variable at its 99th percentile value.

30

equipment and computer equipment; and electronic and other electrical equipment and compo-

nents), and low-tech firms are the rest. This is intended to check whether our results are driven

by a subset of firms and whether they are differential between these two subsamples. The results

are fairly similar in these two subsamples, except for the superstar fraction variable, which shows

a considerably stronger relationship for the low-tech sample.

5.3 Panel Results

As noted above, our baseline (balanced) sample uses one observation per firm and focuses on 1995-

2000. In this section, we investigate several additional issues. First, we show that our results

hold if we look at a considerably larger sample spanning a longer time period. Second, and more

importantly, we also show that, though naturally much noisier, the results are also consistent when

we exploit within-firm variation in the age of the CEO. Third, and consistent with these within-firm

results, we provide some evidence that it is the age of the current CEO that seems to matter most

for the creativity of innovations. Fourth, we also use these results to shed some preliminary light on

the relative importance of the impact of manager age on creative innovations vs. sorting of young

managers across different types of firms.

With this objective in mind, in Table 8 we start with our baseline balanced sample, but now we

compute our measures of creativity of innovations at an annual frequency. The covariates we use

are also at an annual frequency and include a full set of year dummies. In Panel A, we maintain our

key right-hand-side variable, average CEO age over the sample period, which is thus held constant

across years in this panel. In this table, standard errors are robust for arbitrary heteroscedasticity

at the firm level (thus allowing for arbitrary dependence across the observations for the same

firm). These specifications are directly comparable to those in Table 5, and indeed, the coeffi cient

estimates and standard errors are very similar (though they are not identical since the covariates

are now time-varying).21

Panel B turns to an unbalanced panel and extends our sample in two different ways. First,

we include firms that were left out of the balanced panel (i.e., firms for which CEO age or patent

information is available in some but not all years). Second, with the unbalanced panel, we can now

consider a longer sample spanning 1992-2002 (we cannot go before 1992 because of the lack of data

on manager age, and we prefer not to go beyond 2002, as this would make the citation window too

short and thus our measures much less reliable). The resulting sample has 6074 observations (or

5268 observations with tail innovation, since we lose firm-years when no patent is above the median

of the citation distribution). Despite the increase in the number of firms to 1208 (from 279) and

21The number of observations is now lower in columns 3 and 4 because not all firms have patents with citationsabove the median (for tail innovations) or with positive citations (for generality) in all years.

31

the addition of four more years of data, the results are remarkably similar to those in Panel A and

to our baseline estimates.

Panel C allows CEO age to vary across years but also includes firm fixed effects as well as

year effects (and, of course, in this case, SIC industry dummies and firm age are dropped). This

effectively means that the CEO age variable is being identified from changes in CEOs.22 Hence,

this is a very demanding specification investigating whether in years where a firm has a younger

CEO, it tends to have more creative innovations, and this motivates our choice of focusing on the

1992-2002 sample for this exercise. In addition to throwing away all of the (potentially useful)

between-firm variation, another challenge to finding meaningful results in this specification is that

patent applications in one year are often the result of research and product selection from several

past years.23 Though these considerations stack the cards against finding a significant relationship

between CEO age and creative innovations, the results are generally quite consistent with our cross-

sectional estimates from the balanced panel. All of the coeffi cient estimates in these within-firm

regressions, except generality, have the same sign and are statistically significant as in our baseline

results in Table 5. For innovation quality, the magnitude of the estimate is about 18% smaller

than the specification without fixed effects in Panel B (e.g., −0.163 vs. −0.200), and for superstar

fraction and tail innovations, it is about 40% of the magnitude in Panel B.

One concern is that the current CEO may have only a limited impact on innovations patented in

a given year that naturally build on research that had been done many years prior. Counteracting

this is that the current CEOmay set the strategy that determines which innovations are pursued and

marketed and seek patent protection during his or her reign. A natural question is therefore whether

it is current CEO age or lagged CEO age that is more important for the creativity of innovations.

This is investigated in Panel D. Perhaps again somewhat surprisingly, these results show that it is

current CEO age that is the key correlate of our measures of creativity of innovations (except for

generality where we could not find any clear pattern). In fact, in all three specifications (innovation

quality, superstar fraction and tail innovations), current CEO age is statistically significant with a

magnitude close to that in Panel C, while lagged CEO age is not.

A related concern is whether we are partially capturing the persistent effects of past innovations.

22This specification is related to Bertrand and Schoar’s famous (2003) paper on the effect of managers on corporatepolicies, though their sample includes chief financial and operating offi cers as well as lower-level executives andpresidents in contrast to our focus on CEOs.Observe that in our model a high-type firm will pursue an incremental innovation strategy for a while and then

switch to a radical innovation strategy while simultaneously changing its manager to a younger one. In this case, thefixed effect estimator may provide an upper bound on the impact of a younger manager on creative innovations forhigh-type firms.23Recall, however, that patents are classified according to their year of application, so we are investigating the

impact of CEO age not on patents granted when the CEO is in charge but on patents applied for when the CEO isin charge.

32

We investigate this issue by including the lagged dependent variable on the right-hand side. Though

such a model, with fixed effects and lagged dependent variable, is not consistently estimated by the

standard within estimator when the coeffi cient on the lagged dependent variable is close to 1, the

results in Panel E show that its coeffi cient is very far from 1 and the estimates are very similar to

those in Panel C.24

It is noteworthy that the inclusion of firm fixed effects reduces the coeffi cient estimate on CEO

age but only slightly (e.g., for innovation quality, from −0.200 in Panel B to −0.163 in Panel C).

This suggests that part of the relationship we observe in the cross-section is due to the sorting of

younger CEOs to firms that are already more likely to engage in creative innovations, but there

is also a significant impact of a younger CEO on creative innovations for the same firm. This is

also confirmed by the relative explanatory powers of CEO age and firm fixed effects. If we just

include log employment, log sales, log patents, firm age, four-digit SIC dummies and application

year dummies (that is, the same variables as in our baseline regressions except CEO age), the

R2 of the regression for innovation quality is 0.49. Once we add CEO age, this increases to 0.64.

When we also add firm fixed effects, it further increases to 0.77. This suggests that the explanatory

powers of CEO age and firm fixed effects are roughly comparable,25 even though firm fixed effects

here capture not just the sorting channel, so this comparison would provide an upper bound on

the quantitative magnitudes of the sorting channel vs. the direct impact of CEO age on creative

innovations. The pattern is similar for the other measures of creative innovations. Overall, we

tentatively conclude that both the sorting effects and the direct effect of manager age on creativity

of innovations are present and sizable.

In the next subsection, we pursue another approach to shed more light on the effect of manager

age on the creativity of innovations.

5.4 Inventor Age and Creativity of Innovations

We next turn to patent-level regressions to investigate the relationship between the age of inventor–

defined as any inventor listed in our patent data– and our various measures of creativity of innova-

tions. Though in our theoretical model there is no distinction between managers and inventors, this

distinction is of course important in practice. One might then expect the role of product-line man-

agers in our model to be played partly by the top management of the firm and partly by inventors

(or the lead inventor) working on a particular R&D project. CEOs, then, not only decide which

projects the company should focus on but also choose the research team. In this subsection, we24 If we estimate these models using Arellano and Bond’s (1991) GMM estimators, the results are similar with

innovation quality and superstar fraction, but weaker with the tail innovation index, partly because we lose about aquarter of our sample with these GMM models.25Of course, the order in which these two sets of variables are added is important. If firm fixed effects are added

without CEO age, then the R2 is 0.72.

33

bring in information on the age of inventors in order to investigate the effect of manager/inventor

age on the creativity of innovations once we control for the type of characteristics of the firm.

We use Lai et. al.’s (2009) unique inventor identifiers described above to create a proxy for this

variable. Our proxy is the number of years since the first innovation of the inventor, which we will

refer to as “inventor age.”

Our main regression in this subsection will be at the patent level and take the form

yift = φIift + αmft +X′iftβ + δf + γi + dt + εift. (18)

Here yift is one of our measures of the creativity of innovation for (patent) i granted to firm f

at time t. Our key right-hand-side variable is Iift, the age of the inventors named in patent i (in

practice, there are often more than one such inventor listed for a patent). In addition, mft is defined

as CEO age at time t and will be included in some regressions, Xift is a vector of possible controls,

and δf denotes a full set of firm fixed effects, so that our specifications here exploit differences in

the creativity of innovations of a single firm as a function of the characteristics of the innovators

involved in the relevant patent. In our core specifications, we also control for a set of dummies,

denoted by γi, related to inventor characteristics as we described below. All specifications also

control for a full set of year effects, denoted by dt, and εift is the error term.26

The results from the estimation of (18) are reported in Table 9. In Panel A we focus on a

specification similar to the regressions with firm fixed effects reported in Table 8. This is useful

for showing that this different frame still replicates the results showing the impact of CEO age

on creativity of innovations. In particular, Panel A focuses on Compustat firms for the period

1992-2002 and includes the same set of controls as in Table 8 Panel C (firm fixed effects, year fixed

effects, log employment, log sales and log patents of the firm); it does not contain any variables

related to inventor characteristics. As in the rest of this table, these regressions are not weighted

(since they are at the patent level) and the standard errors are robust and clustered at the firm

level.

Our results using this specification are similar to those of Panel C of Table 8, though a little

smaller. In column 1, for instance, we see an estimate of −0.108 (standard error = 0.040) compared

to −0.163 in Table 8. We cannot define our measure of the superstar fraction and tail innovations

in these patent-level regressions. We can, however, look at a patent-level measure related to tail

innovations, a dummy for the patent in question being above the pth percentile of the citation

distribution. We report results using this measure for two values, p = 0.99 and p = 0.90, in

columns 2 and 3. Both of these measures are negatively correlated with CEO age, though only

26A single patent can appear multiple times in our sample if it belongs to multiple firms, but this is very rare andapplies to less than 0.2% of the patents in our sample.

34

marginally significantly in these specifications. Finally, we also show results with the generality

index, even though the results in Table 8 already indicated that, with firm fixed effects included,

there is no longer a significant relationship between CEO age and the generality index, and this

lack of relationship persists for all of the estimates we report in Table 9 (and for this reason, though

we do show them for completeness, we will not discuss them in detail).

Panel B goes in the other direction and reports the estimates of a model that controls for

inventor characteristics and looks at the impact of inventor age, without controlling for CEO age,

for the same sample as in Panel A (thus restricting it to firms with information on CEO age). As

with all of the other models reported in this table, in Panel B we control for a full set of dummies

for the maximum number of patents of any inventor associated with the patent in question has

over our sample period;27 a full set of dummies for the size of the inventor team (i.e., how many

inventors are listed); and a full set of dummies for the three-digit IPC class.28 The inclusion of

this rich set of dummy variables enables us to compare inventors of similar productivity. It thus

approximates a model that includes a full set of inventor dummies.29 The results show that there

is a strong relationship between inventor age and the creativity of innovations. For example, in

column 1, the coeffi cient estimate on inventor age is −0.236 (standard error = 0.027), about twice

as large as the CEO age estimate in Panel A.

When we do not control for CEO age, the sample can be extended beyond 1992-2002. This is

done in Panel C, which expands the sample in two different ways, first by including Compustat

firms without CEO information, and second by broadening the time period covered to 1985-2002.

The results are very similar to those in Panel B, indicating that the focus on Compustat firms with

CEO age information is not responsible for the broad patterns we are documenting.

Panel D extends the sample further to non-Compustat firms, which can also be included in our

analysis since we are not using information on CEO age. This increases our sample sixfold (since

most patents are held by non-Compustat firms). However, in this case, we can no longer include

the employment and sales controls. Despite the addition of almost 1.5 million additional patents

and the lack of our employment and sales controls, the results in this panel are again very similar

to those in previous panels, and suggest that, at least in this instance, our results are not driven

by our focus on the Compustat sample.

Panel E provides our main results in this subsection. It returns to the Compustat sample over

27 In other words, we include a dummy variable for the assignee/inventor of this patent with the highest numberof total patents having k = 1, 2, ..., 89+ patents (where 89+ corresponds to 89 or more patents for the inventor withthe maximum number of patents).28This corresponds to 374 separate technology classes and is roughly at the same level of disaggregation as the 375

SIC dummies we used in the firm-level analysis in Tables 5-7.29We cannot include a full set of inventor fixed effects directly because inventor age would not be identified in this

case since we also have a full set of year dummies.

35

the period 1992-2002 and adds back the CEO age variable; otherwise, the specification is identical

to that in Panel B. The results show precisely estimated impacts of both CEO age and inventor

age. For example, in column 1 with our innovation quality variable, the coeffi cient on CEO age

is −0.111 (standard error = 0.038) and that on inventor age is −0.235 (standard error = 0.027);

these are very close to the estimates in Panels A and B, respectively. The pattern is similar in the

other columns (except again for generality).

These results provide further evidence that manager/CEO age has an effect on the creativity

of innovations that goes beyond the sorting of managers to firms. The quantitative magnitudes are

smaller than those in our baseline firm-level regressions, which suggests that some of the firm-level

relationship was indeed capturing the endogenous sorting of younger managers toward firms that

are more open to disruption. Our next results, reported in Table 10, provide some direct evidence

on this by looking at the relationship between inventor age and CEO age. In particular, we estimate

a regression similar to equation (18) except that now the dependent variable is the average age of

the inventors on the patents granted for that firm in year t and the key right-hand-side variable is

the age of the CEO, and firm fixed effects are again controlled for. The first column of Table 10

reports a regression of the average age of inventors on firm and year fixed effects, log employment,

log sales, log patents, and CEO age, while the second column also adds dummies for inventor team

size and three-digit IPC class as in the specifications in Table 9. The results, which show a positive

(though only marginally significant) relationship, suggest that younger CEOs tend to hire younger

inventors, indirectly corroborating the sorting effect emphasized in our theoretical model.30

5.5 Stock of Knowledge, Opportunity Cost and Creativity of Innovations

Finally, Table 11 turns to an investigation of some additional implications of our approach already

highlighted in our theoretical model (in particular, Proposition 5). We noted there that we may

expect openness to disruption to be more important for companies that are technologically more

advanced (as measured by the number of patents), but also that companies that have more to

lose (because of the greater opportunity cost of disruption in terms of other profitable activities)

may shy away from disruptive creative innovations. The firm-level data enable us to look at this

issue by including the interaction between CEO age and log total patent count (as a proxy for how

advanced the technology of the company is) and also the interaction between CEO age and log

sales (as a proxy for company revenues that may be risked by disruptive innovations). According

to the theoretical ideas suggested above, we expect the interaction with log total patent count to

be negative, and that with sales to be positive (indicating that average manager age matters more

for the creativity of innovations for companies with a significant number of patents and less for

30 Interestingly, this result disappears when we do not control for firm fixed effects.

36

companies with high sales).

This is a demanding, as well as crude, test, since neither proxy is perfect, and moreover, log

sales and log patent counts are positively correlated (the weighted correlation between the two

variables is 0.7 in our sample), thus stacking the cards against finding an informative set of results.

Nevertheless, Table 11, which uses the same unbalanced sample with annual observations as in

Table 8 Panel C, provides some evidence that our theoretical expectations are partially borne out.

In all of our specifications, the interaction between CEO age and log total patent count is negative

and the interaction with log sales is positive. Moreover, these interactions are statistically significant

except for the log patent interaction for the innovation quality measure.31 These results thus provide

some support for the hypothesis that the stock of knowledge of the company and opportunity cost

effects might be present and might in fact be quite important (at least quantitatively at this

correlational level).

6 Conclusion

Despite a large theoretical and now a growing empirical literature on innovation, there is relatively

little work on the determinants of the creativity of innovative activity, and in particular, the like-

lihood of innovations and patents that contribute most to knowledge. In this paper, building on

Schumpeter’s ideas, we suggested that openness to new ideas, disruptive innovations and uncon-

ventional practices– which we called openness to disruption, for short– may be a key determinant

of creative innovations, and likewise, resistance to such disruptive behavior may hold back some of

the most creative innovative activities.

We provided a simple model drawing a clear distinction between radical (more creative) innova-

tions and incremental innovations, whereby the former combines ideas from several different lines of

research and creates more significant advances (and contributions to knowledge). These advances

can be discouraged or even stopped, either through pecuniary or non-pecuniary means preventing

radical innovations directly or discouraging cross-fertilization of ideas from different fields.

The bulk of our paper provides illustrative cross-country and firm-level correlations consistent

with the role of openness to disruption. We use several measures to proxy for creative innovations.

These include our proxy for innovation quality, which is the average number of citations per patent;

two indices for creativity of innovations, which are the fraction of superstar innovators and the like-

lihood of a very high number of citations (in particular, tail citations relative to median citations);

and the generality index.

Our main proxy for openness to disruption is the age of the CEO or top management of the

31As noted above, the main effects are evaluated at the sample mean and are typically close to the estimatesreported in Table 5.

37

company (or the average age of the CEO and CFO of the top 25 publicly listed companies in a

country). This variable is motivated based on the idea that only companies or societies open to

such disruption will allow the young to rise up within the hierarchy. This is the only variable we

have available as a proxy for openness to disruption at the firm level. At the country level, we

augment this variable with the popular indices for individualism and uncertainty avoidance based

on the work by the Dutch social scientist Geert Hofstede. Using these proxies, at the country, firm

and patent level, we find fairly consistent and robust correlations between openness to disruption

and creative innovations. We also show that these relationships are generally robust.

Our theoretical model also suggests that the impact of openness to disruption should be larger

for companies that are technologically more advanced (closer to the technology frontier) and smaller

for companies that have a greater opportunity cost of disruptive innovation. We reported results

from our firm-level data confirming this pattern as well.

We view our paper as a first step in the study of the impact of various social and economic

incentives on creative activities and, in particular, on creative innovations. Future work investi-

gating the causal effect of various other firm-level or cross-country characteristics on the creativity

of innovations is a natural direction, which could both rely on instrumental-variables strategies

and exploit the structural relationships implied by a more detailed microeconomic model. Another

fruitful direction would be to systematically investigate what types of firms and firm organizations

encourage creativity and lead to more creative innovations. This would involve both theoretical and

empirical analyses of the internal organization of firms and their research strategies and a study of

the interplay between institutional and society-level factors and the internal organization of firms.

Appendix

Proof of Proposition 2. We conjecture that the value function for low-type firms takes the

form in (10) . Substituting this conjecture into (7), we get

r [Aqj +Bqtαn]−Bgqtαn = πqj + ξ

[Aqj +Aηqtα

n+1 +Bqtαn+1

−Aqj −Bqtαn]

−τAqj − τBqtαn

Equating the coeffi cients on qj and qtαn, we obtain

rA = π − τA

and

rB −Bg = αξAη + ξB (α− 1)− τB.

Solving these equations for A and B completes the proof.

38

Proof of Proposition 3. Following the same steps, we conjecture that the value function for

high-type firms takes the form in (12), and substitute this into (8) to get

r[Aqj + qtB (n)

]− gqtB (n) = max

πqj + ξ[Aqtηα

n+1 + qtB (n+ 1)− qtB (n)]

;

πqj + Λqa∗θH

[Aqt + Aηqt + qtB (0)

] −τ[Aqj + qtB (n)

]which implies

(r + τ)[Aqj + qtB (n)

]− gqtB (n) = πqj + max

qtξ[Aηαn+1 + B (n+ 1)− B (n)

];

Λqa∗θH

[Aqt + Aηqt + qtB (0)

] Once again equating coeffi cients, we obtain

A =π

r + τ

and

(r − g + τ) B (n) = max

ξ[Aηαn+1 + B (n+ 1)− B (n)

];

Λqa∗θH

[(1 + η) A+ B (0)

] Now define B (n) implicitly as

(r − g + τ) B (n) = ξ[Aηαn+1 + B (n+ 1)− B (n)

]This function can be written as

B (n) = βAηαn+1 + βB (n+ 1)

where β = ξ(r−g+τ+ξ) . From standard dynamic programming arguments (e.g., Theorem 4.7 in Stokey

and Lucas, 1989), B (n) is strictly decreasing.

This implies that there exists n∗ such that firms with n < n∗ will undertake incremental

innovation and will switch to radical innovation at n∗. The expression for n∗ follows by equating

the value of pursuing radical and incremental innovations at n′ and setting n∗ as the smallest integer

greater than n′.

Proof of Proposition 5. Define q∗n,t as the quality level that makes a cluster that has had

n prior innovations so far at time t just indifferent between radical incremental innovation (for a

high-type firm). Then we have that

(r+τ)VH(q∗n,t, n

)−VH

(q∗n,t, n

)= max

{πq∗n,t + ξ

[VH(q∗n,t + ηn+1,t(q

∗n,t), n+ 1

)− VH

(q∗n,t, n

)];

πq∗n,t + ΛqaθHEVH(t)

},

39

where we have written explicitly ηn+1,t(q) as the incremental improvement in productivity starting

from quality q that has been improved n times already and average quality in the economy is qt

(subsumed in the time argument t). Clearly ηn+1,t(q) is increasing in q. Therefore:

VH(q∗n,t + ηn+1,t(q

∗n,t), n+ 1

)− VH

(q∗n,t, n

)=

Λqaθ

ξEVH(t) (19)

and

(r + τ)VH(q∗n,t, n

)− VH

(q∗n,t, n

)= πq∗n,t + ΛqaθEVH(t) (20)

Now we will consider two alternative cases:

Case 1:

q∗n+1,t ≥ q∗n,t + ηn+1,t(q∗n,t). (21)

This condition implies that if a particular high-type firm finds it optimal to switch to radical

innovation today, but instead undertakes a successful incremental innovation (by mistake or off-the-

equilibrium path), then subsequently it will still want to immediately switch to radical innovation.

Under this case, we have

(r+τ)VH(q∗n,t + ηn+1,t(q

∗n,t), n+ 1

)−VH

(q∗n,t + ηn+1,t(q

∗n,t), n+ 1

)= πq∗n,t+πηn+1,t(q

∗n,t)+ΛqaθEVH(t).

(22)

This follows from the fact that, by definition, in this case, at q∗n,t + ηn+1,t(q∗n,t), the firm will want

to switch to radical innovation.

Now differentiating (19) with respect to time, we have

VH(q∗n,t + ηn+1,t(q

∗n,t), n+ 1

)− VH

(q∗n,t, n

)=

Λqaθ

ξ∂EVH(t)/∂t =

Λqaθ

ξgEVH(t),

where we have used the fact that in a stationary equilibrium EVH(t) grows at the rate g. We can

now subtract (20) from (22) to obtain:

(r + τ)[VH(q∗n,t + ηn+1,t(q

∗n,t), n+ 1

)− VH

(q∗n,t, n

)] = πηn+1,t(q

∗n,t) +

Λqaθ

ξgEVH(t) (23)

Then, combining (19) and (23) we can derive

ηn+1,t(q∗n,t) =

r − g + τ

πξΛqaθEVH(t). (24)

In this case, for all q less than q∗n,t, it is optimal to switch to radical innovation. Moreover, q∗n,t is

increasing in n. We next derive the condition under which (21) indeed applies.

Defining vt ≡ r−g+τπξ ΛqaaθEVH(t), equation (24) implies:

[κqt + (1− κ)q∗n,t]ηαn+1 = v(t),

40

or

q∗n,t =v(t)/ηαn+1 − κqt

1− κ , (25)

and similarly

q∗n+1,t =v(t)/ηαn+2 − κqt

1− κ . (26)

Combining these equations, we obtain that (21) is satisfied if

(1− κ)ηαn+2 + α ≤ 1. (27)

Case 2:

q∗n+1,t − ηn+1,t(q∗n,t) < q∗n,t. (28)

This implies that if a high-type firm is indifferent between radical and incremental innovation at

n + 1st prior incremental innovations at time t, then it would have preferred to switch to radical

innovation at nth prior incremental innovations. This condition is clearly the complement of (21).

In this case, start with q∗n+1,t, which satisfies (22). Under condition (28), q∗n,t satisfies (20), so

we again arrive at (19) and (24). This gives the same expression as for q∗n,t and q∗n+1,t as in (25)

and (26). Thus the condition for (28) to be satisfied, with an identical argument, is

(1− κ)ηαn+2 + α > 1,

which is the complement of (27), and thus establishes that (25) still applies, and thus for all q less

than q∗n,t, it is optimal to switch to radical innovation, and q∗n,t is increasing in n. This completes

the proof of the proposition.

41

References

Abrams, D. S., U. Akcigit, and J. Popadak (2013). Patent Value and Citations: Creative

Destruction or Defensive Disruption? National Bureau of Economic Research WP 19763.

Acemoglu, D., U. Akcigit, N. Bloom, and W. R. Kerr (2013). Innovation, Reallocation and

Growth. National Bureau of Economic Research WP 18993.

Acs, Z. J. and D. B. Audretsch (1987). Innovation, Market Structure, and Firm Size. Review

of Economics and Statistics, 567-574.

Acs, Z. J. and D. B. Audretsch (1988). Innovation in Large and Small Firms: An Empirical

Analysis. American Economic Review, 678-690.

Adner, R. and P. Zemsky (2005). Disruptive Technologies and the Emergence of Competition.

RAND Journal of Economics, 229-254.

Adorno, T. W., E. Frenkel-Brunswik, D. J. Levinson, and R. N. Sanford (1950). The Authori-

tarian Personality. Harpers.

Akcigit, U. and W. R. Kerr (2010). Growth through Heterogeneous Innovations. National

Bureau of Economic Research WP 16443.

Alesina, A., P. Giuliano, and N. Nunn (2011). Fertility and the Plough. American Economic

Review, 101 (3), 499-503.

Arellano, M. and S. Bond (1991). Some Tests of Specification for Panel Data: Monte Carlo

Evidence and an Application to Employment Equations. Review of Economic Studies, 58(2), 277-

297.

Arrow, K. (1962). Economic Welfare and the Allocation of Resources for Invention. NBER

Chapters, 609-626.

Azoulay, P., G. Manso and J. G. Zivin (2011). Incentives and Creativity: Evidence from the

Academic Life Sciences. RAND Journal of Economics 42(3), 527-544.

Azoulay, P., J. G. Zivin, and J. Wang (2010). Superstar Extinction. Quarterly Journal of

Economics. 125(2), 549-589.

Baumol, W. J. (2009). Small Enterprises, Large Firms and Growth. In The Role of SMEs

and Entrepreneurship in a Globalised Economy, ed. by A. Lundström, Ministry of Education and

Research Sweden, 59 —72.

Barro, R. and J. W. Lee (forthcoming). A New Data Set of Educational Attainment in the

World, 1950-2010. Journal of Development Economics.

Bertrand, M. and A. Schoar (2003). Managing with Style: The Effect of Managers on Firm

Policies. Quarterly Journal of Economics, 118(4), 1169-1208.

Christensen, C. (1997). The Innovator’s Dilemma: When New Technologies Cause Great Firms

42

to Fail. Harvard Business Press.

Cyert, R. M. and G. March, James (1963). A Behavioral Theory of the Firm. Englewood Cliffs,

NJ: Prentice-Hall.

Durkheim, E. (1933). The Division of Labor. New York: Macmillan.

Fernandez, R. and A. Fogli (2009). Culture: An Empirical Investigation of Beliefs, Work and

Fertility. American Economic Journal: Macroeconomics, 1 (1), 146-177.

Fogli A. and L Veldkamp (2013). Germs, Social Networks and Growth. National Bureau of

Economic Research WP 18470.

Galenson D. and B. Weinberg (1999). Age and the Quality of Work: The Case of Modern

American Painters. National Bureau of Economic Research WP 7122.

Galenson D. and B. Weinberg (2001). Creating Modern Art: The Changing Careers of Painters

in France From Impressionism to Cubism. American Economic Review, 91 (4), 1063-1071.

Gennaioli, N. and R. La Porta, F. Lopez-de-Silanes, and A. Shleifer (2013). Human Capital

and Regional Development, Quarterly Journal of Economics, 128(1), 105—164.

Gorodnichenko Y. and G. Roland (2012). Culture, Institutions and the Wealth of Nations. UC

Berkeley Working Paper.

Greif, A. (1994). Cultural Beliefs and the Organization of Society: A Historical and Theoretical

Reflection on Collectivist and Individualist Societies. Journal of Political Economy, 102(5), 912-

950.

Guiso, L., P. Sapienza, and L. Zingales (2010). Civic Capital as the Missing Link. National

Bureau of Economic Research WP 15845.

Hall, B. H., A. B. Jaffe, and M. Trajtenberg (2001). The NBER Patent Citation Data File:

Lessons, Insights and Methodological Tools. National Bureau of Economic Research WP 8498.

Harhoff, D., F. Narin, F. M. Scherer, and K. Vopel (1999). Citation Frequency and the Value

of Patented Inventions. Review of Economics and Statistics, 81(3), 511-515.

Henderson, R. (2006). The Innovator’s Dilemma as a Problem of Organizational Competence.

Journal of Product Innovation Management, 23 (1), 5-11.

Hofstede, G., G. J. Hofstede, and M. Minkov (2010). Cultures and Organizations: Software for

the Mind. McGraw-Hill.

Hofstede, G. H. (2001). Culture’s Consequences: Comparing Values, Behaviors, Institutions

and Organizations across Nations. Sage.

Jones, B. F. (2009). The Burden of Knowledge and the Death of the Renaissance Man: Is

Innovation Getting Harder? Review of Economic Studies, 76 (1), 283-317.

Jones, B. F. (2010). Age and Great Invention. Review of Economics and Statistics, 92 (1),

1-14.

43

Jones B. F. and B. Weinberg (2011). A Dynamics and Scientific Creativity. Proceedings of the

National Academy of Sciences. 108 (47), 18910-18914.

King, A. A. and C. L. Tucci (2002). Incumbent Entry into New Market Niches: The Role of

Experience and Managerial Choice in the Creation of Dynamic Capabilities. Management Science,

48 (2), 171-186.

Kirzner, I. M. (1997). Entrepreneurial Discovery and the Competitive Market Process: An

Austrian Approach. Journal of Economic Literature, 35 (1), 60-85.

Klette, T. J. and S. Kortum (2004). Innovating Firms and Aggregate Innovation. Journal of

Political Economy, 112, 986-1018.

Kluckhohn, F. R. and F. L. Strodtbeck (1961). Variations in Value Orientations. Evanston,

IL: Row Peterson.

Kortum, S. and J. Lerner (2000). Assessing the Contribution of Venture Capital to Innovation.

RAND Journal of Economics, 674-692.

Lai, R., A. D’Amour, A. Yu, Y. Sun, and L. Fleming (2009). Disambiguation and Co-author

Networks of the US Patent Inventor Database. Harvard Business School Working Paper.

McClelland, D. (1961). The Achieving Society. The Free Press, New York.

Parsons, T. (1949). Essays in Sociological Theory; Pure and Applied. Free Press.

Sampat, B. N. and A. A. Ziedonis (2004). Patent Citations and the Economic Value of Patents.

In Handbook of Quantitative Science and Technology Research, 277-298. Springer.

Sarada, and Tocoian , O. (2013). The Success of Entrepreneurial Networks: Evidence from

Brazil. Duke University Working Paper.

Schumpeter, J. A. (1934). The Theory of Economic Development: An Inquiry into Profits,

Capital, Credit, Interest, and the Business Cycle. University of Illinois at Urbana-Champaign’s

Academy for Entrepreneurial Leadership Historical Research Reference in Entrepreneurship.

Schwartz, S. H. (1994). Beyond Individualism/Collectivism: New Cultural Dimensions of Val-

ues. Sage Publications, Inc.

Schwartz, S. H. and W. Bilsky (1990). Toward a Theory of the Universal Content and Struc-

ture of Values: Extensions and Cross-cultural Replications. Journal of Personality and Social

Psychology, 58(5), 878.

Shane, H. and M. Klock (1997). The Relation between Patent Citations and Tobin’s Q in the

Semiconductor Industry. Review of Quantitative Finance and Accounting, 9 (2), 131-146.

Shane, S. (1993). Cultural Influences on National Rates of Innovation. Journal of Business

Venturing, 8(1), 59-73.

Shane, S. (1995). Is the Independent Entrepreneurial Firm a Valuable Organizational Form?

In Academy of Management Proceedings, 110-113. Academy of Management.

44

Stokey, N. and R. Lucas Jr. (1989) Recursive Methods in Economic Dynamics, Harvard Uni-

versity Press.

Tabellini, G. (2008a). Institutions and Culture. Journal of European Economic Association,

6(2), 255-294.

Tabellini, G. (2008b). The Scope of Cooperation: Values and Incentives. Quarterly Journal of

Economics, 123(3), 905-950.

Trajtenberg, M. (1990). A Penny for Your Quotes: Patent Citations and the Value of Innova-

tions. RAND Journal of Economics, 172-187.

Triandis, H. C. (1995). Individualism & Collectivism. Westview Press.

Weinberg B. and D. Galenson (2005). Creative Careers: Lifecycles of Noble Laureates in

Economics. National Bureau of Economic Research WP 11799.

Weitzman, M. L. (1998). Recombinant Growth. Quarterly Journal of Economics, 113(2), 331-

360.

Winslow, E. K. and G. T. Solomon (1987). Entrepreneurs Are More Than Non-Conformists:

They Are Mildly Sociopathic. Journal of Creative Behavior, 21(3), 202-213.

45

Table 1: Summary Statistics

Panel A: Descriptive Statistics

Variable Observations Mean Standard Deviation

Cross-Country Sample (Country Averages, 1995-2000)

individualism 50 .813 .263uncertainty aversion 50 .492 .195average manager age 37 56.1 2.98innovation quality 50 14.5 3.26superstar fraction 50 6.68 3.65tail innovation 50 1.92 .945generality 50 21.0 1.81log patents 50 10.5 1.52log income per capita 50 10.3 .305secondary years of schooling 50 4.84 .827R&D intensity 44 2.59 .363

Balanced Firm Sample (Firm Averages, 1995-2000)

CEO age 279 55.3 6.47average manager age 279 52.3 4.32innovation quality 279 20.5 8.76superstar fraction 279 12.3 10.1tail innovation 279 2.72 2.56generality 279 21.5 5.53log patents 279 5.86 1.51log employment 279 3.84 1.38log sale 279 4.34 1.47firm age 279 37.3 14.4R&D intensity 257 8.52 17.0

Unbalanced Firm Sample (Annual Firm Observations, 1992-2002)

CEO age 6074 55.5 6.83average manager age 6074 52.3 4.43innovation quality 6074 17.3 10.6superstar fraction 6074 10.5 11.0tail innovation 5268 2.79 3.78generality 5697 19.9 9.23log patents 6074 5.66 1.60log employment 6074 3.73 1.50log sale 6074 4.15 1.60firm age 6074 34.9 16.0

- Table 1 continued on next page -

Panel B: Correlation Matrix of Openness to Disruption Variables

individualism uncertainty aversion average manager ageindividualism 1.000uncertainty aversion -0.884 1.000average manager age -0.770 0.844 1.000

Panel C: Correlation Matrix of Cross-Country Innovation Variables

innovation quality superstar fraction tail innovation generalityinnovation quality 1.000superstar fraction 0.932 1.000tail innovation 0.945 0.990 1.000generality 0.902 0.880 0.906 1.000

Panel D: Correlation Matrix of Firm-Level Innovation Variables

innovation quality superstar fraction tail innovation generalityinnovation quality 1.000superstar fraction 0.925 1.000tail innovation 0.893 0.829 1.000generality -0.177 -0.204 -0.145 1.000

Notes: All statistics in this table are weighted by the number of patents (of the country or the firm). Individualism and

uncertainty aversion are Hofstede’s indices of national cultures (and are normalized to lie between 0 and 1), and country

average manager age is the average manager of CEOs and CFOs of up to the 25 largest firms in the country. Innovation

quality is the average number of citations per patent (using the truncation correction weights devised by Hall, Jaffe, and

Trajtenberg, 2001); superstar fraction is the fraction of patents accounted for by superstar researchers (those above the

95th percentile of the citation distribution); tail innovation is the fraction of patents of a country or firm above the 99th

percentile of the citation distribution divided by the fraction of patents above the median of the distribution; and generality

index measures the dispersion of citations received across two-digit IPC technology classes. Log income per capita at the

country level, and log employment, log sales at the firm level are computed as the average of, respectively, annual log

income per capita, log employment and log sale between 1995 and 2000. CEO age is the age of the CEO and average

manager age is the average age of the top management, both from the Execucomp dataset. The balanced firm panel is the

sample of firms from Compustat with complete data on CEO age, employment, sales, and firm age and positive patents in

each year between 1995 and 2000. The unbalanced firm panel is a sample of firms from Compustat with at least one year

of complete data between 1992 and 2002. See text for the definition of other variables and further details.

Table 2: Baseline Cross-Country Regressions

Innovation Quality Superstar Fraction Tail Innovation Generality

Panel A: Individualism

individualism 4.965 9.929 2.369 3.420(2.461) (2.393) (0.640) (0.487)

log income per capita -1.233 -2.130 -0.472 -0.252(1.195) (1.270) (0.334) (0.373)

secondary years of schooling -0.467 -0.317 -0.056 -0.051(1.229) (1.174) (0.323) (0.227)

log patents 1.622 1.125 0.308 0.725(0.490) (0.472) (0.129) (0.164)

R2 0.73 0.81 0.79 0.83N 50 50 50 50

Panel B: Uncertainty Avoidance

uncertainty aversion -8.354 -13.528 -3.174 -4.242(2.946) (2.715) (0.722) (0.798)

log income per capita -0.408 -0.657 -0.124 0.232(0.957) (0.600) (0.177) (0.558)

secondary years of schooling -0.745 -0.346 -0.054 0.008(1.149) (1.108) (0.307) (0.208)

log patent 1.708 1.257 0.339 0.765(0.439) (0.424) (0.125) (0.189)

R2 0.80 0.86 0.84 0.84N 50 50 50 50

Panel C: Average Manager Age

manager age -0.484 -0.960 -0.225 -0.278(0.225) (0.221) (0.058) (0.056)

log income per capita -0.491 -0.702 -0.136 0.211(1.153) (1.066) (0.291) (0.468)

secondary years of schooling -1.000 -1.359 -0.291 -0.231(1.481) (1.462) (0.396) (0.341)

log patent 2.232 2.331 0.591 1.072(0.706) (0.695) (0.193) (0.222)

R2 0.74 0.82 0.80 0.80N 37 37 37 37

Notes: Weighted cross-country regressions with total number of patents as weights. The dependent variables are innovation

quality, superstar fraction, tail innovation, and generality (the last three are multiplied by 100 to ease legibility). See

text and notes to Table 1 for variable definitions. Each country observation is the sample average between 1995-2000 as

described in the text and the notes to Table 1. Robust standard errors are in parentheses.

Table 3: Cross-Country Regressions (Alternative Measures)

Innovation Quality Superstar Fraction Tail Innovation Originality

(5 years) (Best Patent) (90/50)

Panel A: Individualism

individualism 2.039 0.052 9.966 8.015(1.009) (0.045) (4.028) (0.653)

R2 0.74 0.80 0.68 0.91N 50 50 50 50

Panel B: Uncertainty Avoidance

uncertainty aversion -3.461 -0.106 -15.964 -9.084(1.215) (0.057) (4.689) (1.336)

R2 0.81 0.83 0.78 0.87N 50 50 50 50

Panel C: Average Manager Age

manager age -0.203 -0.005 -1.002 -0.713(0.092) (0.004) (0.372) (0.083)

R2 0.75 0.80 0.70 0.88N 37 37 37 37

Notes: Weighted cross-country regressions with total number of patents as weights. The dependent variables are alternative

measures of innovation quality (computed over the next five years), superstar fraction (with superstars defined according to the

best patent), tail innovation (with fraction of patents above the 90th percentile of the citation distribution in the numerator), and

the originality index (the last three are multiplied by 100 to ease legibility). Each regression also controls for log income per capita,

average years of secondary schooling, and log total patents. See text and notes to Table 1 for variable definitions. Each country

observation is the sample average between 1995-2000 as described in the text and the notes to Table 1. Robust standard errors are

in parentheses.

Table 4: Cross-Country Regressions (Controlling for R&D Intensity)

Innovation Quality Superstar Fraction Tail Innovation Generality

Panel A: Individualism

individualism 8.245 13.786 3.291 2.932(2.821) (2.602) (0.725) (0.778)

R2 0.78 0.85 0.83 0.83N 44 44 44 44

Panel B: Uncertainty Avoidance

uncertainty aversion -9.589 -14.173 -3.305 -3.452(2.747) (2.753) (0.754) (0.915)

R2 0.82 0.86 0.83 0.85N 44 44 44 44

Panel C: Average Manager Age

manager age -0.636 -1.096 -0.257 -0.622(0.255) (0.253) (0.066) (0.105)

R2 0.76 0.83 0.81 0.91N 33 33 33 33

Notes: Weighted cross-country regressions with total number of patents as weights. The dependent variables are innovation

quality, superstar fraction, tail innovation, and generality (the last three are multiplied by 100 to ease legibility). Each regression

also controls for log income per capita, average years of secondary schooling, log total patents, and R&D intensity defined

as total R&D expenditure divided by GDP. See text and notes to Table 1 for variable definitions. Each country observation

is the sample average between 1995-2000 as described in the text and the notes to Table 1. Robust standard errors are in parentheses.

Table 5: Baseline Firm-Level Regressions

Innovation Quality Superstar Fraction Tail Innovation Generality

CEO age -0.278 -0.300 -0.151 -0.183(0.088) (0.141) (0.054) (0.055)

firm age -0.219 -0.238 -0.063 0.029(0.078) (0.106) (0.029) (0.046)

log employment -1.599 -4.813 -0.908 -4.574(1.937) (3.376) (0.793) (1.500)

log sales 1.833 5.215 0.743 4.421(1.425) (2.645) (0.650) (1.331)

log patent 1.073 0.093 0.662 -0.696(0.769) (1.336) (0.356) (0.633)

R2 0.88 0.81 0.79 0.83N 279 279 279 279

Notes: Weighted cross-sectional regressions with total number of patents as weights. The sample is the balanced firm

panel and each observation is the sample average between 1995-2000 as described in the notes to Table 1. The dependent

variables are innovation quality, superstar fraction, tail innovation, and generality (the last three are multiplied by 100 to

ease legibility). In addition, all regressions control for a full set of dummies for four-digit SIC industries. See text and notes

to Table 1 for variable definitions. Robust standard errors are in parentheses.

Table 6: Firm-Level Regressions (Alternative Measures)

Innovation Quality Superstar Fraction Tail Innovation Originality(5 years) (Best Patent) (90/50)

CEO age -0.129 -0.497 -0.299 -0.285(0.041) (0.332) (0.094) (0.075)

R2 0.87 0.87 0.83 0.87N 279 279 279 279

Notes: Weighted cross-sectional regressions with total number of patents as weights. The sample is the balanced firm panel

and each observation is the sample average between 1995-2000 as described in the notes to Table 1. The dependent variables

are alternative measures of innovation quality (computed over the next five years), superstar fraction (with superstars

defined according to the best patent), tail innovation (with share of the patents of the firm among all the patents above the

90th percentile of the citation distribution in the numerator), and the originality index (the last three are multiplied by 100

to ease legibility). All regressions control for firm age, log employment, log sales, log total patents, and a full set of dummies

for four-digit SIC industries. See text and notes to Table 1 for variable definitions. Robust standard errors are in parentheses.

Table 7: Firm-Level Regressions (Robustness)

Innovation Quality Superstar Fraction Tail Innovation Generality

Panel A: With SIC3 Dummies

CEO age -0.257 -0.284 -0.126 -0.086(0.070) (0.123) (0.050) (0.091)

R2 0.77 0.72 0.64 0.70N 279 279 279 279

Panel B: With Additional Controls

CEO age -0.270 -0.282 -0.150 -0.194(0.090) (0.140) (0.052) (0.054)

R2 0.88 0.82 0.79 0.83N 279 279 279 279

Panel C: With Additional Controls Plus R&D Intensity

CEO age -0.258 -0.295 -0.142 -0.184(0.088) (0.149) (0.048) (0.053)

R2 0.89 0.82 0.81 0.84N 257 257 257 257

Panel D: With Average Manager Age

average manager age -0.418 -0.467 -0.224 -0.339(0.163) (0.206) (0.094) (0.084)

R2 0.87 0.81 0.77 0.83N 279 279 279 279

Panel E: High-Tech Subsample

CEO age -0.227 -0.191 -0.145 -0.189(0.068) (0.157) (0.045) (0.043)

R2 0.92 0.84 0.86 0.81N 87 87 87 87

Panel F: Low-Tech Subsample

CEO age -0.439 -0.704 -0.143 -0.153(0.200) (0.252) (0.085) (0.146)

R2 0.85 0.82 0.72 0.86N 192 192 192 192

Notes: Weighted cross-sectional regressions with total number of patents as weights. The sample is the balanced firm panel

and each of the ratios is the sample average 1995-2000 as described in the notes to Table 1.The dependent variables are

innovation quality, superstar fraction, tail innovation, and generality (the last three are multiplied by 100 to ease legibility).

Each panel is for a different specification. Unless otherwise stated, all regressions control for firm age, log employment,

log sales, log total patents, and four-digit SIC dummies (see text and notes to Table 1 for variable definitions). Robust

standard errors are in parentheses. Panel A controls for three-digit SIC dummies instead of the four-digit dummies. Panel

B adds to the specification of Table 5 profitability (profit over sales), indebtedness (debt over sales) and log physical capital.

Panel C adds to the specification of Panel B R&D intensity (R&D expenditure over sales). Panel D uses average manager

age instead of CEO age. Panels E and F are for the high-tech and low-tech subsamples. High-tech sample includes all

firms with a primary industry classification of SIC 35 (industrial and commercial machinery and equipment and computer

equipment) and 36 (electronic and other electrical equipment and components), while the low-tech sample includes the rest.

Table 8: Firm-Level Panel Regressions

Innovation Quality Superstar Fraction Tail Innovation Generality

Panel A: Average CEO Age (No Fixed Effects), Balanced Firm Sample, 1995-2000

average CEO age -0.227 -0.336 -0.132 -0.183(0.068) (0.103) (0.041) (0.044)

R2 0.70 0.69 0.47 0.75N 1,674 1,674 1,594 1,655

Panel B: Average CEO Age (No Fixed Effects), Unbalanced Firm Sample, 1992-2002

average CEO age -0.200 -0.326 -0.129 -0.162(0.077) (0.130) (0.042) (0.047)

R2 0.63 0.55 0.30 0.74N 6,074 6,074 5,268 5,697

Panel C: CEO Age (Fixed Effects), Unbalanced Firm Sample, 1992-2002

CEO age -0.163 -0.130 -0.044 0.031(0.050) (0.054) (0.021) (0.033)

R2 0.77 0.81 0.52 0.84N 6,074 6,074 5,268 5,697

Panel D: CEO Age and Lagged CEO Age (Fixed Effects), Unbalanced Firm Sample, 1993-2002

CEO age -0.137 -0.104 -0.035 0.028(0.042) (0.041) (0.021) (0.030)

lagged CEO age -0.091 -0.074 -0.020 0.017(0.068) (0.064) (0.030) (0.040)

R2 0.79 0.82 0.54 0.84N 4,858 4,858 4,292 4,562

Panel E: CEO Age and Lagged Dependent Var (Fixed Effects), Unbalanced Firm Sample, 1993-2002

CEO age -0.103 -0.079 -0.035 0.035(0.034) (0.037) (0.017) (0.029)

lagged dependent variable 0.417 0.407 0.162 0.147(0.037) (0.047) (0.047) (0.038)

R2 0.84 0.86 0.55 0.84N 5,522 5,522 4,510 5,091

Notes: Weighted firm-level panel regressions with annual observations with number of patents (in that year) as weights. The dependent variables

are innovation quality, superstar fraction, tail innovation, and generality (the last three are multiplied by 100 to ease legibility). Robust standard

errors clustered at the firm level are in parentheses. Panel A is for our balanced firm sample 1995-2000, and controls for firm age, log employment,

log sales, log patents, a full set of four-digit SIC dummies, and year dummies (and thus no firm dummies), and the key right-hand side variable is

average CEO age (constant over time). Panel B is identical to Panel A except that the sample is extended to the unbalanced firm panel 1992-2002.

In Panel C, the key right-hand side variable is CEO age (in that year), and the regression also includes a full set of firm fixed effects (and thus

firm age and the four-digit SIC dummies are no longer included). Panel D is identical to Panel C except that it also includes a one year lag of

CEO age as well as current CEO age, and Panel E is identical to Panel C except that it also includes a one year lag of the dependent variable on

the right-hand side. See text and notes to Table 1 for variable definitions.

Table 9: Patent-Level Panel Regressions

Innovation Quality Tail Innovation Tail Innovation Generality

(Above 99) (Above 90)

Panel A: CEO Age, Unbalanced Firm Sample, 1992-2002

CEO age -0.108 -0.293 -1.066 0.029(0.040) (0.132) (0.422) (0.026)

R2 0.11 0.03 0.07 0.11N 311,216 311,216 311,216 261,266

Panel B: Inventor Age, Unbalanced Firm Sample, 1992-2002

inventor age -0.236 -0.444 -2.924 -0.019(0.027) (0.120) (0.327) (0.022)

R2 0.14 0.03 0.09 0.15N 311,216 311,216 311,216 261,266

Panel C: Inventor Age, Extended Sample, 1985-2002

inventor age -0.228 -0.376 -2.872 -0.017(0.022) (0.075) (0.296) (0.017)

R2 0.16 0.05 0.10 0.14N 562,552 562,552 562,552 462,313

Panel D: Inventor Age, Extended Sample, 1985-2002

inventor age -0.201 -0.327 -2.359 -0.046(0.010) (0.036) (0.134) (0.011)

R2 0.27 0.15 0.19 0.25N 1,855,887 1,855,887 1,855,887 1,550,825

Panel E: CEO Age and Inventor Age, Unbalanced Firm Sample, 1992-2002

inventor age -0.235 -0.443 -2.919 -0.019(0.027) (0.120) (0.327) (0.022)

CEO age -0.111 -0.302 -1.071 0.029(0.038) (0.126) (0.402) (0.023)

R2 0.14 0.03 0.09 0.15N 311,216 311,216 311,216 261,266

Notes: Patent-level panel regressions with annual observations. The dependent variables are innovation quality at the

patent level; a dummy for the patent being above the 99th percentile of the citation distribution; dummy for the patent

being above the 90th percentile of the citation distribution; and generality index at the patent level (the last three are

multiplied by 100 to ease legibility). Robust standard errors clustered at the firm level are in parentheses. Panel A is for

our unbalanced firm sample 1992-2002 and controls for log employment, log sales, log patents, a full set of firm fixed effects,

and application year dummies, and the key right-and side variable is CEO age. Panel B is for our unbalanced firm sample

1992-2002 and controls for log employment, log sales, log patents, application year dummies, a full set of firm fixed effects,

a full set of dummies for inventor team size, a full set of dummies for three-digit IPC technology class dummies, and a

full set of dummies for the total number of patents of the inventor within the team with the highest number of patents,

and the key right-and side variable is average inventor age. Panel C expands the sample of Panel B to 1985-2002 and

also adds Compustat firms without CEO information into the sample. Panel D extends the sample of Panel C to include

non-Compustat firms as well (hence excludes log sales and log employment). Panel E is for our unbalanced firm sample

1992-2002 and adds CEO age to the specification of Panel B. See text and notes to Table 1 for variable definitions.

Table 10: Inventor Age and CEO Age,Unbalanced Firm Sample, 1992-2002

Inventor age Inventor age

(1) (2)

CEO age 0.012 0.012(0.006) (0.006)

R2 0.11 0.13N 311,216 311,216

Notes: Patent-level panel regressions with annual observations for the unbalanced firm sample 1992-2002. The dependent

variable is the average age of inventors. The first column controls for log employment, log sales, log patents, applica-

tion year dummies, and a full set of firm dummies, and the second column adds to this a full set of team size dummies

and a full set of dummies for three-digit IPC technology class dummies. See text and notes to Table 1 for variable definitions.

Table 11: Stock of Knowledge, Opportunity Cost, and Creative Innovations,Unbalanced Firm Sample, 1992-2002

Innovation Quality Superstar Fraction Tail Innovation Generality

CEO age -0.178 -0.207 -0.073 -0.049(0.030) (0.031) (0.014) (0.018)

log sales 1.477 2.188 0.224 1.309(0.485) (0.672) (0.219) (0.352)

log patent -0.439 -0.268 0.240 -0.051(0.208) (0.284) (0.107) (0.158)

average CEO age × log patent -0.003 -0.067 -0.028 -0.036(0.015) (0.023) (0.009) (0.012)

average CEO age × log sales 0.036 0.086 0.032 0.045(0.019) (0.025) (0.010) (0.012)

R2 0.64 0.55 0.31 0.74N 6,074 6,074 5,268 5,697

Notes: Weighted firm-level panel regressions with annual observations for the unbalanced firm panel, 1992-2002, with

number of patents (in that year) as weights. The dependent variables are innovation quality, superstar fraction, tail

innovation, and generality (the last three are multiplied by 100 to ease legibility). Robust standard errors clustered at

the firm level are in parentheses. All regressions also include log employment, application year dummies and a full set of

dummies for four-digit SIC industries. See text and notes to Table 1 for variable definitions.

Table A1: Average Annual Patent Counts by Country, 1995-2000

Country Abbreviation Patent Count Country Abbreviation Patent Count

Argentina AR 9.2 India IN 90.3

Austria AT 365.0 Italy IT 1439.8

Australia AU 744.0 Japan JP 33954.8

Belgium BE 522.8 South Korea KR 3581.5

Bulgaria BG 3.8 Luxemburg LU 62.8

Brazil BR 69.7 Malta MT 2.0

Canada CA 2433.2 Mexico MX 59.2

Switzerland CH 1588.7 Malaysia MY 14.5

Chile CL 8.8 Netherlands NL 1236.7

China CN 109.5 Norway NO 239.2

Colombia CO 2.0 New Zealand NZ 104.7

Czech Republic CZ 17.7 Poland PL 10.0

Germany DE 9257.0 Portugal PT 8.7

Denmark DK 448.5 Romania RO 2.7

Spain ES 193.8 Russia RU 88.2

Finland FI 910.3 Saudi Arabia SA 18.2

France FR 3877.5 Sweden SE 1691.3

Great Britain GB 2869.5 Singapore SG 191.2

Greece GR 15.7 Slovenia SI 13.7

Hong Kong HK 171.8 Slovakia SK 4.0

Croatia HR 7.7 Thailand TH 10.7

Hungary HU 33.3 Turkey TR 5.3

Indonesia ID 3.0 United States US 93722.5

Ireland IE 111.3 Venezuela VE 24.3

Israel IL 580.7 South Africa ZA 88.7

Notes: This table shows the average annual patent counts between 1995-2000, registered at the USPTO from that country.

Figure 1: Openness to Disruption Measures vs Innovation Quality

AR

AT

AU

BE

BG

BR

CA

CH

CL

CN

CO

CZ

DE

DKES

FI

FR

GB

GR

HK

HR

HU

ID

IEIL

IN

ITJP

KR

LU

MT

MX

MY

NL

NO

NZ

PL

PT

RO

RU

SA

SE

SG

SI

SK

TH

TR

US

VE

ZA

-50

510

15av

erag

e nu

mbe

r of

cita

tions

-.4 0 .4 .8individualism

A. Individualism vs Innovation Quality

AR

AT

AU

BE

BG

BR

CA

CH

CL

CN

CO

CZ

DE

DKES

FI

FR

GB

GR

HK

HR

HU

ID

IEIL

IN

ITJP

KR

LU

MT

MX

MY

NL

NO

NZ

PL

PT

RO

RU

SA

SE

SG

SI

SK

TH

TR

US

VE

ZA

-50

510

15av

erag

e nu

mbe

r of

cita

tions

-.7 -.3 .1 .5uncertainty avoidance

B. Uncertainty Avoidance vs Innovation Quality

AR

AT

AU

BEBR

CA

CH

CN

CZ

DE

DKES

FI

FR

GB

HK

HU

IEIL

IN

ITJP

KR

LU

MX

MY

NL

NO

NZ

PL

RU

SA

SE

SG

US

VE

ZA

-50

510

15av

erag

e nu

mbe

r of

cita

tions

-7 -3 1 5average age

C. Manager Age vs Innovation Quality

Notes: Residual plots from a weighted regression of innovation quality (average number of citations per patents) on Hofstede’s individualism index,

Hofstede’s uncertainty avoidance index, and our average manager age variable on log income per capita, averages of secondary years of schooling, and

log total number of patents, with total number of patents as weights for 1995-2000. See text and notes to Table 1 for variable definitions.

1